C++ Neural Networks & Fuzzy Logic - Valluru B Rao

C++ Neural Networks and Fuzzy Logic:Preface
C++ Neural Networks and Fuzzy Logic

by Valluru B. Rao
MTBooks, IDG Books Worldwide, Inc.
ISBN: 1558515526 Pub Date: 06/01/95
Table of Contents
Preface
The number of models available in neural network literature is quite large. Very often the treatment is
mathematical and complex. This book provides illustrative examples in C++ that the reader can use as a basis
for further experimentation. A key to learning about neural networks to appreciate their inner workings is to
experiment. Neural networks, in the end, are fun to learn about and discover. Although the language for
description used is C++, you will not find extensive class libraries in this book. With the exception of the
backpropagation simulator, you will find fairly simple example programs for many different neural network
architectures and paradigms. Since backpropagation is widely used and also easy to tame, a simulator is
provided with the capacity to handle large input data sets. You use the simulator in one of the chapters in this
book to solve a financial forecasting problem. You will find ample room to expand and experiment with the
code presented in this book.
There are many different angles to neural networks and fuzzy logic. The fields are expanding rapidly with
ever−new results and applications. This book presents many of the different neural network topologies,
including the BAM, the Perceptron, Hopfield memory, ART1, Kohonen’s Self−Organizing map, Kosko’s
Fuzzy Associative memory, and, of course, the Feedforward Backpropagation network (aka Multilayer
Perceptron). You should get a fairly broad picture of neural networks and fuzzy logic with this book. At the
same time, you will have real code that shows you example usage of the models, to solidify your
understanding. This is especially useful for the more complicated neural network architectures like the
Adaptive Resonance Theory of Stephen Grossberg (ART).
The subjects are covered as follows:
• Chapter 1 gives you an overview of neural network terminology and nomenclature. You discover
that neural nets are capable of solving complex problems with parallel computational architectures.
The Hopfield network and feedforward network are introduced in this chapter.
• Chapter 2 introduces C++ and object orientation. You learn the benefits of object−oriented
programming and its basic concepts.
• Chapter 3 introduces fuzzy logic, a technology that is fairly synergistic with neural network
problem solving. You learn about math with fuzzy sets as well as how you can build a simple
fuzzifier in C++.
• Chapter 4 introduces you to two of the simplest, yet very representative, models of: the Hopfield
network, the Perceptron network, and their C++ implementations.
• Chapter 5 is a survey of neural network models. This chapter describes the features of several
models, describes threshold functions, and develops concepts in neural networks.
• Chapter 6 focuses on learning and training paradigms. It introduces the concepts of supervised
and unsupervised learning, self−organization and topics including backpropagation of errors, radial
basis function networks, and conjugate gradient methods.
• Chapter 7 goes through the construction of a backpropagation simulator. You will find this
simulator useful in later chapters also. C++ classes and features are detailed in this chapter.
• Chapter 8 covers the Bidirectional Associative memories for associating pairs of patterns.
Preface 1
• Chapter 9 introduces Fuzzy Associative memories for associating pairs of fuzzy sets.
• Chapter 10 covers the Adaptive Resonance Theory of Grossberg. You will have a chance to
experiment with a program that illustrates the working of this theory.
• Chapters 11 and 12 discuss the Self−Organizing map of Teuvo Kohonen and its application to
pattern recognition.
• Chapter 13 continues the discussion of the backpropagation simulator, with enhancements made
to the simulator to include momentum and noise during training.
• Chapter 14 applies backpropagation to the problem of financial forecasting, discusses setting up a
backpropagation network with 15 input variables and 200 test cases to run a simulation. The problem
is approached via a systematic 12−step approach for preprocessing data and setting up the problem.
You will find a number of examples of financial forecasting highlighted from the literature. A
resource guide for neural networks in finance is included for people who would like more information
about this area.
• Chapter 15 deals with nonlinear optimization with a thorough discussion of the Traveling
Salesperson problem. You learn the formulation by Hopfield and the approach of Kohonen.
• Chapter 16 treats two application areas of fuzzy logic: fuzzy control systems and fuzzy databases.
This chapter also expands on fuzzy relations and fuzzy set theory with several examples.
• Chapter 17 discusses some of the latest applications using neural networks and fuzzy logic.
In this second edition, we have followed readers’ suggestions and included more explanations and material, as
well as updated the material with the latest information and research. We have also corrected errors and
omissions from the first edition.
Neural networks are now a subject of interest to professionals in many fields, and also a tool for many areas of
problem solving. The applications are widespread in recent years, and the fruits of these applications are being
reaped by many from diverse fields. This methodology has become an alternative to modeling of some
physical and nonphysical systems with scientific or mathematical basis, and also to expert systems
methodology. One of the reasons for it is that absence of full information is not as big a problem in neural
networks as it is in the other methodologies mentioned earlier. The results are sometimes astounding, even
phenomenal, with neural networks, and the effort is at times relatively modest to achieve such results. Image
processing, vision, financial market analysis, and optimization are among the many areas of application of
neural networks. To think that the modeling of neural networks is one of modeling a system that attempts to
mimic human learning is somewhat exciting. Neural networks can learn in an unsupervised learning mode.
Just as human brains can be trained to master some situations, neural networks can be trained to recognize
patterns and to do optimization and other tasks.
In the early days of interest in neural networks, the researchers were mainly biologists and psychologists.
Serious research now is done by not only biologists and psychologists, but by professionals from computer
science, electrical engineering, computer engineering, mathematics, and physics as well. The latter have either
joined forces, or are doing independent research parallel with the former, who opened up a new and promising
field for everyone.
In this book, we aim to introduce the subject of neural networks as directly and simply as possible for an easy
understanding of the methodology. Most of the important neural network architectures are covered, and we
earnestly hope that our efforts have succeeded in presenting this subject matter in a clear and useful fashion.
We welcome your comments and suggestions for this book, from errors and oversights, to suggestions for
improvements to future printings at the following E−mail addresses:
V. Rao rao@cse.bridgeport.edu
Preface 2
H. Rao ViaSW@aol.com
Table of Contents
Copyright © IDG Books Worldwide, Inc.
Preface 3

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Preface
Dedication
Chapter 1—Introduction to Neural Networks

Neural Processing
Neural Network
Output of a Neuron
Cash Register Game
Weights
Training
Feedback
Supervised or Unsupervised Learning
Noise
Memory
Capsule of History
Neural Network Construction
Sample Applications
Qualifying for a Mortgage
Cooperation and Competition
Example—A Feed−Forward Network
Example—A Hopfield Network
Hamming Distance
Asynchronous Update
Binary and Bipolar Inputs
Bias
Another Example for the Hopfield Network
Summary
Chapter 2—C++ and Object Orientation

Introduction to C++
Encapsulation
Data Hiding
Constructors and Destructors as Special Functions of C++
Dynamic Memory Allocation
Overloading
Polymorphism and Polymorphic Functions
Overloading Operators
Inheritance
Derived Classes
Reuse of Code
C++ Compilers
Writing C++ Programs
Summary
Preface 4
Chapter 3—A Look at Fuzzy Logic

Crisp or Fuzzy Logic?
Fuzzy Sets
Fuzzy Set Operations
Union of Fuzzy Sets
Intersection and Complement of Two Fuzzy Sets
Applications of Fuzzy Logic
Examples of Fuzzy Logic
Commercial Applications
Fuzziness in Neural Networks
Code for the Fuzzifier
Fuzzy Control Systems
Neural−Trained Fuzzy Systems
Summary
Chapter 4—Constructing a Neural Network

First Example for C++ Implementation
Classes in C++ Implementation
C++ Program for a Hopfield Network
Header File for C++ Program for Hopfield Network
Notes on the Header File Hop.h
Source Code for the Hopfield Network
Comments on the C++ Program for Hopfield Network
Output from the C++ Program for Hopfield Network
Further Comments on the Program and Its Output
A New Weight Matrix to Recall More Patterns
Weight Determination
Binary to Bipolar Mapping
Pattern’s Contribution to Weight
Autoassociative Network
Orthogonal Bit Patterns
Network Nodes and Input Patterns
Second Example for C++ Implementation
C++ Implementation of Perceptron Network
Header File
Implementation of Functions
Source Code for Perceptron Network
Comments on Your C++ Program
Input/Output for percept.cpp
Network Modeling
Tic−Tac−Toe Anyone?
Stability and Plasticity
Stability for a Neural Network
Plasticity for a Neural Network
Short−Term Memory and Long−Term Memory
Summary
Chapter 5—A Survey of Neural Network Models
Preface 5
Neural Network Models

Layers in a Neural Network
Single−Layer Network
XOR Function and the Perceptron
Linear Separability
A Second Look at the XOR Function: Multilayer Perceptron
Example of the Cube Revisited
Strategy
Details
Performance of the Perceptron
Other Two−layer Networks
Many Layer Networks
Connections Between Layers
Instar and Outstar
Weights on Connections
Initialization of Weights
A Small Example
Initializing Weights for Autoassociative Networks
Weight Initialization for Heteroassociative Networks
On Center, Off Surround
Inputs
Outputs
The Threshold Function
The Sigmoid Function
The Step Function
The Ramp Function
Linear Function
Applications
Some Neural Network Models
Adaline and Madaline
Backpropagation
Figure for Backpropagation Network
Bidirectional Associative Memory
Temporal Associative Memory
Brain−State−in−a−Box
Counterpropagation
Neocognitron
Adaptive Resonance Theory
Summary
Chapter 6—Learning and Training

Objective of Learning
Learning and Training
Hebb’s Rule
Delta Rule
Supervised Learning
Generalized Delta Rule
Statistical Training and Simulated Annealing
Radial Basis−Function Networks
Unsupervised Networks
Preface 6
Self−Organization
Learning Vector Quantizer
Associative Memory Models and One−Shot Learning
Learning and Resonance
Learning and Stability
Training and Convergence
Lyapunov Function
Other Training Issues
Adaptation
Generalization Ability
Summary
Chapter 7—Backpropagation
Feedforward Backpropagation Network
Mapping
Layout
Training
Illustration: Adjustment of Weights of Connections from a Neuron in the Hidden Layer
Illustration: Adjustment of Weights of Connections from a Neuron in the Input Layer
Adjustments to Threshold Values or Biases
Another Example of Backpropagation Calculations
Notation and Equations
Notation
Equations
C++ Implementation of a Backpropagation Simulator
A Brief Tour of How to Use the Simulator
C++ Classes and Class Hierarchy
Summary
Chapter 8—BAM: Bidirectional Associative Memory

Introduction
Inputs and Outputs
Weights and Training
Example
Recall of Vectors
Continuation of Example
Special Case—Complements
C++ Implementation
Program Details and Flow
Program Example for BAM
Header File
Source File
Program Output
Additional Issues
Unipolar Binary Bidirectional Associative Memory
Summary
Chapter 9—FAM: Fuzzy Associative Memory

Introduction
Association
Preface 7
FAM Neural Network

Encoding
Example of Encoding
Recall
C++ Implementation
Program details
Header File
Source File
Output
Summary
Chapter 10—Adaptive Resonance Theory (ART)

Introduction
The Network for ART1
A Simplified Diagram of Network Layout
Processing in ART1
Special Features of the ART1 Model
Notation for ART1 Calculations
Algorithm for ART1 Calculations
Initialization of Parameters
Equations for ART1 Computations
Other Models
C++ Implementation
A Header File for the C++ Program for the ART1 Model Network
A Source File for C++ Program for an ART1 Model Network
Program Output
Summary
Chapter 11—The Kohonen Self−Organizing Map

Introduction
Competitive Learning
Normalization of a Vector
Lateral Inhibition
The Mexican Hat Function
Training Law for the Kohonen Map
Significance of the Training Law
The Neighborhood Size and Alpha
C++ Code for Implementing a Kohonen Map
The Kohonen Network
Modeling Lateral Inhibition and Excitation
Classes to be Used
Revisiting the Layer Class
A New Layer Class for a Kohonen Layer
Implementation of the Kohonen Layer and Kohonen Network
Flow of the Program and the main() Function
Flow of the Program
Results from Running the Kohonen Program
A Simple First Example
Orthogonal Input Vectors Example
Variations and Applications of Kohonen Networks
Preface 8
Using a Conscience
LVQ: Learning Vector Quantizer
Counterpropagation Network
Application to Speech Recognition
Summary
Chapter 12—Application to Pattern Recognition

Using the Kohonen Feature Map
An Example Problem: Character Recognition
C++ Code Development
Changes to the Kohonen Program
Testing the Program
Generalization versus Memorization
Adding Characters
Other Experiments to Try
Summary
Chapter 13—Backpropagation II
Enhancing the Simulator
Another Example of Using Backpropagation
Adding the Momentum Term
Code Changes
Adding Noise During Training
One Other Change—Starting Training from a Saved Weight File
Trying the Noise and Momentum Features
Variations of the Backpropagation Algorithm
Applications
Summary
Chapter 14—Application to Financial Forecasting

Introduction
Who Trades with Neural Networks?
Developing a Forecasting Model
The Target and the Timeframe
Domain Expertise
Gather the Data
Pre processing the Data for the Network
Reduce Dimensionality
Eliminate Correlated Inputs Where Possible
Design a Network Architecture
The Train/Test/Redesign Loop
Forecasting the S&P 500
Choosing the Right Outputs and Objective
Choosing the Right Inputs
Choosing a Network Architecture
Preprocessing Data
A View of the Raw Data
Highlight Features in the Data
Normalizing the Range
The Target
Preface 9
Storing Data in Different Files

Training and Testing
Using the Simulator to Calculate Error
Only the Beginning
What’s Next?
Technical Analysis and Neural Network Preprocessing
Moving Averages
Momentum and Rate of Change
Relative Strength Index
Percentage R
Herrick Payoff Index
MACD
“Stochastics”
On−Balance Volume
Accumulation−Distribution
What Others Have Reported
Can a Three−Year−Old Trade Commodities?
Forecasting Treasury Bill and Treasury Note Yields
Neural Nets versus Box−Jenkins Time−Series Forecasting
Neural Nets versus Regression Analysis
Hierarchical Neural Network
The Walk−Forward Methodology of Market Prediction
Dual Confirmation Trading System
A Turning Point Predictor
The S&P 500 and Sunspot Predictions
A Critique of Neural Network Time−Series Forecasting for Trading
Resource Guide for Neural Networks and Fuzzy Logic in Finance
Magazines
Books
Book Vendors
Consultants
Historical Financial Data Vendors
Preprocessing Tools for Neural Network Development
Genetic Algorithms Tool Vendors
Fuzzy Logic Tool Vendors
Neural Network Development Tool Vendors
Summary
Chapter 15—Application to Nonlinear Optimization

Introduction
Neural Networks for Optimization Problems
Traveling Salesperson Problem
The TSP in a Nutshell
Solution via Neural Network
Example of a Traveling Salesperson Problem for Hand Calculation
Neural Network for Traveling Salesperson Problem
Network Choice and Layout
Inputs
Activations, Outputs, and Their Updating
Performance of the Hopfield Network
Preface 10
C++ Implementation of the Hopfield Network for the Traveling Salesperson

Problem
Source File for Hopfield Network for Traveling Salesperson Problem
Output from Your C++ Program for the Traveling Salesperson Problem
Other Approaches to Solve the Traveling Salesperson Problem
Optimizing a Stock Portfolio
Tabu Neural Network
Summary
Chapter 16—Applications of Fuzzy Logic

Introduction
A Fuzzy Universe of Applications
Section I: A Look at Fuzzy Databases and Quantification
Databases and Queries
Relations in Databases
Fuzzy Scenarios
Fuzzy Sets Revisited
Fuzzy Relations
Matrix Representation of a Fuzzy Relation
Properties of Fuzzy Relations
Similarity Relations
Resemblance Relations
Fuzzy Partial Order
Fuzzy Queries
Extending Database Models
Example
Possibility Distributions
Example
Queries
Fuzzy Events, Means and Variances
Example: XYZ Company Takeover Price
Probability of a Fuzzy Event
Fuzzy Mean and Fuzzy Variance
Conditional Probability of a Fuzzy Event
Conditional Fuzzy Mean and Fuzzy Variance
Linear Regression a la Possibilities
Fuzzy Numbers
Triangular Fuzzy Number
Linear Possibility Regression Model
Section II: Fuzzy Control
Designing a Fuzzy Logic Controller
Step One: Defining Inputs and Outputs for the FLC
Step Two: Fuzzify the Inputs
Step Three: Set Up Fuzzy Membership Functions for the Output(s)
Step Four: Create a Fuzzy Rule Base
Step Five: Defuzzify the Outputs
Advantages and Disadvantages of Fuzzy Logic Controllers
Summary
Chapter 17—Further Applications
Preface 11
Introduction
Computer Virus Detector
Mobile Robot Navigation
A Classifier
A Two−Stage Network for Radar Pattern Classification
Crisp and Fuzzy Neural Networks for Handwritten Character Recognition
Noise Removal with a Discrete Hopfield Network
Object Identification by Shape
Detecting Skin Cancer
EEG Diagnosis
Time Series Prediction with Recurrent and Nonrecurrent Networks
Security Alarms
Circuit Board Faults
Warranty Claims
Writing Style Recognition
Commercial Optical Character Recognition
ART−EMAP and Object Recognition
Summary
References
Appendix A
Appendix B
Glossary
Index
Preface 12

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Table of Contents
Dedication
To the memory of
Vydehamma, Annapurnamma, Anandarao, Madanagopalarao, Govindarao, and Rajyalakshamma.
Acknowledgments
We thank everyone at MIS:Press/Henry Holt and Co. who has been associated with this project for their
diligence and support, namely, the Technical Editors of this edition and the first edition for their suggestions
and feedback; Laura Lewin, the Editor, and all of the other people at MIS:Press for making the book a reality.
We would also like to thank Dr. Tarek Kaylani for his helpful suggestions, Professor R. Haskell, and our other
readers who wrote to us, and Dr. V. Rao’s students whose suggestions were helpful. Please E−mail us more
feedback!
Finally, thanks to Sarada and Rekha for encouragement and support. Most of all, thanks to Rohini and Pranav
for their patience and understanding through many lost evenings and weekends.
Table of Contents
Dedication 13

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Previous Table of Contents Next
Chapter 1
Introduction to Neural Networks
Neural Processing
How do you recognize a face in a crowd? How does an economist predict the direction of interest rates? Faced
with problems like these, the human brain uses a web of interconnected processing elements called neurons to
process information. Each neuron is autonomous and independent; it does its work asynchronously, that is,
without any synchronization to other events taking place. The two problems posed, namely recognizing a face
and forecasting interest rates, have two important characteristics that distinguish them from other problems:
First, the problems are complex, that is, you can’t devise a simple step−by−step algorithm or precise formula
to give you an answer; and second, the data provided to resolve the problems is equally complex and may be
noisy or incomplete. You could have forgotten your glasses when you’re trying to recognize that face. The
economist may have at his or her disposal thousands of pieces of data that may or may not be relevant to his
or her forecast on the economy and on interest rates.
The vast processing power inherent in biological neural structures has inspired the study of the structure itself
for hints on organizing human−made computing structures. Artificial neural networks, the subject of this
book, covers the way to organize synthetic neurons to solve the same kind of difficult, complex problems in a
similar manner as we think the human brain may. This chapter will give you a sampling of the terms and
nomenclature used to talk about neural networks. These terms will be covered in more depth in the chapters to
follow.
Neural Network
A neural network is a computational structure inspired by the study of biological neural processing. There are
many different types of neural networks, from relatively simple to very complex, just as there are many
theories on how biological neural processing works. We will begin with a discussion of a layered
feed−forward type of neural network and branch out to other paradigms later in this chapter and in other
chapters.
A layered feed−forward neural network has layers, or subgroups of processing elements. A layer of
processing elements makes independent computations on data that it receives and passes the results to another
layer. The next layer may in turn make its independent computations and pass on the results to yet another
layer. Finally, a subgroup of one or more processing elements determines the output from the network. Each
processing element makes its computation based upon a weighted sum of its inputs. The first layer is the input
layer and the last the output layer. The layers that are placed between the first and the last layers are the
hidden layers. The processing elements are seen as units that are similar to the neurons in a human brain, and
hence, they are referred to as cells, neuromimes, or artificial neurons. A threshold function is sometimes used
to qualify the output of a neuron in the output layer. Even though our subject matter deals with artificial
neurons, we will simply refer to them as neurons. Synapses between neurons are referred to as connections,
Chapter 1 Introduction to Neural Networks 14

which are represented by edges of a directed graph in which the nodes are the artificial neurons.
Figure 1.1 is a layered feed−forward neural network. The circular nodes represent neurons. Here there are
three layers, an input layer, a hidden layer, and an output layer. The directed graph mentioned shows the
connections from nodes from a given layer to other nodes in other layers. Throughout this book you will see
many variations on the number and types of layers.
Figure 1.1 A typical neural network.
Output of a Neuron
Basically, the internal activation or raw output of a neuron in a neural network is a weighted sum of its inputs,
but a threshold function is also used to determine the final value, or the output. When the output is 1, the
neuron is said to fire, and when it is 0, the neuron is considered not to have fired. When a threshold function is
used, different results of activations, all in the same interval of values, can cause the same final output value.
This situation helps in the sense that, if precise input causes an activation of 9 and noisy input causes an
activation of 10, then the output works out the same as if noise is filtered out.
To put the description of a neural network in a simple and familiar setting, let us describe an example about a
daytime game show on television, The Price is Right.
Output of a Neuron 15

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Cash Register Game

A contestant in The Price is Right is sometimes asked to play the Cash Register Game. A few products are
described, their prices are unknown to the contestant, and the contestant has to declare how many units of
each item he or she would like to (pretend to) buy. If the total purchase does not exceed the amount specified,
the contestant wins a special prize. After the contestant announces how many items of a particular product he
or she wants, the price of that product is revealed, and it is rung up on the cash register. The contestant must
be careful, in this case, that the total does not exceed some nominal value, to earn the associated prize. We can
now cast the whole operation of this game, in terms of a neural network, called a Perceptron, as follows.
Consider each product on the shelf to be a neuron in the input layer, with its input being the unit price of that
product. The cash register is the single neuron in the output layer. The only connections in the network are
between each of the neurons (products displayed on the shelf) in the input layer and the output neuron (the
cash register). This arrangement is usually referred to as a neuron, the cash register in this case, being an
instar in neural network terminology. The contestant actually determines these connections, because when the
contestant says he or she wants, say five, of a specific product, the contestant is thereby assigning a weight of
5 to the connection between that product and the cash register. The total bill for the purchases by the
contestant is nothing but the weighted sum of the unit prices of the different products offered. For those items
the contestant does not choose to purchase, the implicit weight assigned is 0. The application of the dollar
limit to the bill is just the application of a threshold, except that the threshold value should not be exceeded for
the outcome from this network to favor the contestant, winning him or her a good prize. In a Perceptron, the
way the threshold works is that an output neuron is supposed to fire if its activation value exceeds the
threshold value.
Weights
The weights used on the connections between different layers have much significance in the working of the
neural network and the characterization of a network. The following actions are possible in a neural network:
1. Start with one set of weights and run the network. (NO TRAINING)
2. Start with one set of weights, run the network, and modify some or all the weights, and run the
network again with the new set of weights. Repeat this process until some predetermined goal is met.
(TRAINING)
Training
Since the output(s) may not be what is expected, the weights may need to be altered. Some rule then needs to
be used to determine how to alter the weights. There should also be a criterion to specify when the process of
successive modification of weights ceases. This process of changing the weights, or rather, updating the
weights, is called training. A network in which learning is employed is said to be subjected to training.
Training is an external process or regimen. Learning is the desired process that takes place internal to the
network.
Cash Register Game 16

Feedback
If you wish to train a network so it can recognize or identify some predetermined patterns, or evaluate some
function values for given arguments, it would be important to have information fed back from the output
neurons to neurons in some layer before that, to enable further processing and adjustment of weights on the
connections. Such feedback can be to the input layer or a layer between the input layer and the output layer,
sometimes labeled the hidden layer. What is fed back is usually the error in the output, modified appropriately
according to some useful paradigm. The process of feedback continues through the subsequent cycles of
operation of the neural network and ceases when the training is completed.
Supervised or Unsupervised Learning

A network can be subject to supervised or unsupervised learning. The learning would be supervised if external
criteria are used and matched by the network output, and if not, the learning is unsupervised. This is one broad
way to divide different neural network approaches. Unsupervised approaches are also termed self−organizing.
There is more interaction between neurons, typically with feedback and intralayer connections between
neurons promoting self−organization.
Supervised networks are a little more straightforward to conceptualize than unsupervised networks. You apply
the inputs to the supervised network along with an expected response, much like the Pavlovian conditioned
stimulus and response regimen. You mold the network with stimulus−response pairs. A stock market
forecaster may present economic data (the stimulus) along with metrics of stock market performance (the
response) to the neural network to the present and attempt to predict the future once training is complete.
You provide unsupervised networks with only stimulus. You may, for example, want an unsupervised
network to correctly classify parts from a conveyor belt into part numbers, providing an image of each part to
do the classification (the stimulus). The unsupervised network in this case would act like a look−up memory
that is indexed by its contents, or a Content−Addressable−Memory (CAM).
Feedback 17

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Noise
Noise is perturbation, or a deviation from the actual. A data set used to train a neural network may have
inherent noise in it, or an image may have random speckles in it, for example. The response of the neural
network to noise is an important factor in determining its suitability to a given application. In the process of
training, you may apply a metric to your neural network to see how well the network has learned your training
data. In cases where the metric stabilizes to some meaningful value, whether the value is acceptable to you or
not, you say that the network converges. You may wish to introduce noise intentionally in training to find out
if the network can learn in the presence of noise, and if the network can converge on noisy data.
Memory
Once you train a network on a set of data, suppose you continue training the network with new data. Will the
network forget the intended training on the original set or will it remember? This is another angle that is
approached by some researchers who are interested in preserving a network’s long−term memory (LTM) as
well as its short−term memory (STM). Long−term memory is memory associated with learning that persists
for the long term. Short−term memory is memory associated with a neural network that decays in some time
interval.
Capsule of History
You marvel at the capabilities of the human brain and find its ways of processing information unknown to a
large extent. You find it awesome that very complex situations are discerned at a far greater speed than what a
computer can do.
Warren McCulloch and Walter Pitts formulated in 1943 a model for a nerve cell, a neuron, during their
attempt to build a theory of self−organizing systems. Later, Frank Rosenblatt constructed a Perceptron, an
arrangement of processing elements representing the nerve cells into a network. His network could recognize
simple shapes. It was the advent of different models for different applications.
Those working in the field of artificial intelligence (AI) tried to hypothesize that you can model thought
processes using some symbols and some rules with which you can transform the symbols.
A limitation to the symbolic approach is related to how knowledge is representable. A piece of information is
localized, that is, available at one location, perhaps. It is not distributed over many locations. You can easily
see that distributed knowledge leads to a faster and greater inferential process. Information is less prone to be
damaged or lost when it is distributed than when it is localized. Distributed information processing can be
fault tolerant to some degree, because there are multiple sources of knowledge to apply to a given problem.
Even if one source is cut off or destroyed, other sources may still permit solution to a problem. Further, with
subsequent learning, a solution may be remapped into a new organization of distributed processing elements
that exclude a faulty processing element.
Noise 18
In neural networks, information may impact the activity of more than one neuron. Knowledge is distributed
and lends itself easily to parallel computation. Indeed there are many research activities in the field of
hardware design of neural network processing engines that exploit the parallelism of the neural network
paradigm. Carver Mead, a pioneer in the field, has suggested analog VLSI (very large scale integration)
circuit implementations of neural networks.
Neural Network Construction

There are three aspects to the construction of a neural network:
1. Structure—the architecture and topology of the neural network

2. Encoding—the method of changing weights
3. Recall—the method and capacity to retrieve information
Let’s cover the first one—structure. This relates to how many layers the network should contain, and what
their functions are, such as for input, for output, or for feature extraction. Structure also encompasses how
interconnections are made between neurons in the network, and what their functions are.
The second aspect is encoding. Encoding refers to the paradigm used for the determination of and changing of
weights on the connections between neurons. In the case of the multilayer feed−forward neural network, you
initially can define weights by randomization. Subsequently, in the process of training, you can use the
backpropagation algorithm, which is a means of updating weights starting from the output backwards. When
you have finished training the multilayer feed−forward neural network, you are finished with encoding since
weights do not change after training is completed.
Finally, recall is also an important aspect of a neural network. Recall refers to getting an expected output for a
given input. If the same input as before is presented to the network, the same corresponding output as before
should result. The type of recall can characterize the network as being autoassociative or heteroassociative.
Autoassociation is the phenomenon of associating an input vector with itself as the output, whereas
heteroassociation is that of recalling a related vector given an input vector. You have a fuzzy remembrance of
a phone number. Luckily, you stored it in an autoassociative neural network. When you apply the fuzzy
remembrance, you retrieve the actual phone number. This is a use of autoassociation. Now if you want the
individual’s name associated with a given phone number, that would require heteroassociation. Recall is
closely related to the concepts of STM and LTM introduced earlier.
The three aspects to the construction of a neural network mentioned above essentially distinguish between
different neural networks and are part of their design process.
Neural Network Construction 19


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Sample Applications
One application for a neural network is pattern classification, or pattern matching. The patterns can be
represented by binary digits in the discrete cases, or real numbers representing analog signals in continuous
cases. Pattern classification is a form of establishing an autoassociation or heteroassociation. Recall that
associating different patterns is building the type of association called heteroassociation. If you input a
corrupted or modified pattern A to the neural network, and receive the true pattern A, this is termed
autoassociation. What use does this provide? Remember the example given at the beginning of this chapter. In
the human brain example, say you want to recall a face in a crowd and you have a hazy remembrance (input).
What you want is the actual image. Autoassociation, then, is useful in recognizing or retrieving patterns with
possibly incomplete information as input. What about heteroassociation? Here you associate A with B. Given
A, you get B and sometimes vice versa. You could store the face of a person and retrieve it with the person’s
name, for example. It’s quite common in real circumstances to do the opposite, and sometimes not so well.
You recall the face of a person, but can’t place the name.
Qualifying for a Mortgage
Another sample application, which is in fact in the works by a U.S. government agency, is to devise a neural
network to produce a quick response credit rating of an individual trying to qualify for a mortgage. The
problem to date with the application process for a mortgage has been the staggering amount of paperwork and
filing details required for each application. Once information is gathered, the response time for knowing
whether or not your mortgage is approved has typically taken several weeks. All of this will change. The
proposed neural network system will allow the complete application and approval process to take three hours,
with approval coming in five minutes of entering all of the information required. You enter in the applicant’s
employment history, salary information, credit information, and other factors and apply these to a trained
neural network. The neural network, based on prior training on thousands of case histories, looks for patterns
in the applicant’s profile and then produces a yes or no rating of worthiness to carry a particular mortgage.
Let’s now continue our discussion of factors that distinguish neural network models from each other.
Cooperation and Competition

We will now discuss cooperation and competition. Again we start with an example feed forward neural
network. If the network consists of a single input layer and an output layer consisting of a single neuron, then
the set of weights for the connections between the input layer neurons and the output neuron are given in a
weight vector. For three inputs and one output, this could be W = {w1, w2, w3 }. When the output layer has
more than one neuron, the output is not just one value but is also a vector. In such a situation each neuron in
one layer is connected to each neuron in the next layer, with weights assigned to these interconnections. Then
the weights can all be given together in a two−dimensional weight matrix, which is also sometimes called a
correlation matrix. When there are in−between layers such as a hidden layer or a so−called Kohonen layer or
a Grossberg layer, the interconnections are made between each neuron in one layer and every neuron in the
next layer, and there will be a corresponding correlation matrix. Cooperation or competition or both can be
imparted between network neurons in the same layer, through the choice of the right sign of weights for the
Sample Applications 20
connections. Cooperation is the attempt between neurons in one neuron aiding the prospect of firing by
another. Competition is the attempt between neurons to individually excel with higher output. Inhibition, a
mechanism used in competition, is the attempt between neurons in one neuron decreasing the prospect of
another neuron’s firing. As already stated, the vehicle for these phenomena is the connection weight. For
example, a positive weight is assigned for a connection between one node and a cooperating node in that
layer, while a negative weight is assigned to inhibit a competitor.
To take this idea to the connections between neurons in consecutive layers, we would assign a positive weight
to the connection between one node in one layer and its nearest neighbor node in the next layer, whereas the
connections with distant nodes in the other layer will get negative weights. The negative weights would
indicate competition in some cases and inhibition in others. To make at least some of the discussion and the
concepts a bit clearer, we preview two example neural networks (there will be more discussion of these
networks in the chapters that follow): the feed−forward network and the Hopfield network.
Sample Applications 21

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Example—A Feed−Forward Network

A sample feed−forward network, as shown in Figure 1.2, has five neurons arranged in three layers: two
neurons (labeled x1 and x2) in layer 1, two neurons (labeled x3 and x4) in layer 2, and one neuron (labeled x5)
in layer 3. There are arrows connecting the neurons together. This is the direction of information flow. A
feed−forward network has information flowing forward only. Each arrow that connects neurons has a weight
associated with it (like, w31 for example). You calculate the state, x, of each neuron by summing the weighted
values that flow into a neuron. The state of the neuron is the output value of the neuron and remains the same
until the neuron receives new information on its inputs.
Figure 1.2 A feed−forward neural network with topology 2−2−1.
For example, for x3 and x5:
x3 = w23 x2 + w13 x1
x5 = w35 x3 + w45 x4
We will formalize the equations in Chapter 7, which details one of the training algorithms for the
feed−forward network called Backpropagation.
Note that you present information to this network at the leftmost nodes (layer 1) called the input layer. You
can take information from any other layer in the network, but in most cases do so from the rightmost node(s),
which make up the output layer. Weights are usually determined by a supervised training algorithm, where
you present examples to the network and adjust weights appropriately to achieve a desired response. Once you
have completed training, you can use the network without changing weights, and note the response for inputs
that you apply. Note that a detail not yet shown is a nonlinear scaling function that limits the range of the
weighted sum. This scaling function has the effect of clipping very large values in positive and negative
directions for each neuron so that the cumulative summing that occurs across the network stays within
reasonable bounds. Typical real number ranges for neuron inputs and outputs are –1 to +1 or 0 to +1. You will
see more about this network and applications for it in Chapter 7. Now let us contrast this neural network with
a completely different type of neural network, the Hopfield network, and present some simple applications for
the Hopfield network.
Example—A Hopfield Network

The neural network we present is a Hopfield network, with a single layer. We place, in this layer, four
neurons, each connected to the rest, as shown in Figure 1.3. Some of the connections have a positive weight,
and the rest have a negative weight. The network will be presented with two input patterns, one at a time, and
it is supposed to recall them. The inputs would be binary patterns having in each component a 0 or 1. If two
Example—A Feed−Forward Network 22

patterns of equal length are given and are treated as vectors, their dot product is obtained by first multiplying
corresponding components together and then adding these products. Two vectors are said to be orthogonal, if
their dot product is 0. The mathematics involved in computations done for neural networks include matrix
multiplication, transpose of a matrix, and transpose of a vector. Also see Appendix B. The inputs (which are
stable, stored patterns) to be given should be orthogonal to one another.
Figure 1.3 Layout of a Hopfield network.
The two patterns we want the network to recall are A = (1, 0, 1, 0) and B = (0, 1, 0, 1), which you can verify
to be orthogonal. Recall that two vectors A and B are orthogonal if their dot product is equal to zero. This is
true in this case since
A1B1 + A2 B2 + A3B3 + A4B4 = (1x0 + 0x1 + 1x0 + 0x1) = 0
The following matrix W gives the weights on the connections in the network.
0 −3 3 −3
−3 0 −3 3
W = 3 −3 0 −3
−3 3 −3 0
We need a threshold function also, and we define it as follows. The threshold value [theta] is 0.
1 if t >= [theta]
f(t) = { 0 if t < [theta]
Example—A Feed−Forward Network 23


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
We have four neurons in the only layer in this network. We need to compute the activation of each neuron as
the weighted sum of its inputs. The activation at the first node is the dot product of the input vector and the
first column of the weight matrix (0 −3 3 −3). We get the activation at the other nodes similarly. The output of
a neuron is then calculated by evaluating the threshold function at the activation of the neuron. So if we
present the input vector A, the dot product works out to 3 and f(3) = 1. Similarly, we get the dot products of
the second, third, and fourth nodes to be –6, 3, and –6, respectively. The corresponding outputs therefore are
0, 1, and 0. This means that the output of the network is the vector (1, 0, 1, 0), same as the input pattern. The
network has recalled the pattern as presented, or we can say that pattern A is stable, since the output is equal
to the input. When B is presented, the dot product obtained at the first node is –6 and the output is 0. The
outputs for the rest of the nodes taken together with the output of the first node gives (0, 1, 0, 1), which means
that the network has stable recall for B also.
NOTE: In Chapter 4, a method of determining the weight matrix for the Hopfield network
given a set of input vectors is presented.
So far we have presented easy cases to the network—vectors that the Hopfield network was specifically
designed (through the choice of the weight matrix) to recall. What will the network give as output if we
present a pattern different from both A and B? Let C = (0, 1, 0, 0) be presented to the network. The
activations would be –3, 0, –3, 3, making the outputs 0, 1, 0, 1, which means that B achieves stable recall.
This is quite interesting. Suppose we did intend to input B and we made a slight error and ended up presenting
C, instead. The network did what we wanted and recalled B. But why not A? To answer this, let us ask is C
closer to A or B? How do we compare? We use the distance formula for two four−dimensional points. If (a, b,
c, d) and (e, f, g, h) are two four−dimensional points, the distance between them is:
[radic][(a – e)2 + (b – f)2 + (c – g)2 + (d – h)2 ]
The distance between A and C is [radic]3, whereas the distance between B and C is just 1. So since B is closer
in this sense, B was recalled rather than A. You may verify that if we do the same exercise with D = (0, 0, 1,
0), we will see that the network recalls A, which is closer than B to D.
Hamming Distance
When we talk about closeness of a bit pattern to another bit pattern, the Euclidean distance need not be
considered. Instead, the Hamming distance can be used, which is much easier to determine, since it is the
number of bit positions in which the two patterns being compared differ. Patterns being strings, the Hamming
distance is more appropriate than the Euclidean distance.
NOTE: The weight matrix W we gave in this example is not the only weight matrix that
would enable the network to recall the patterns A and B correctly. You can see that if we
replace each of 3 and –3 in the matrix by say, 2 and –2, respectively, the resulting matrix
would also facilitate the same performance from the network. For more details, consult
Hamming Distance 24
Chapter 4.
Asynchronous Update
The Hopfield network is a recurrent network. This means that outputs from the network are fed back as
inputs. This is not apparent from Figure 1.3, but is clearly seen from Figure 1.4.
Figure 1.4 Feedback in the Hopfield network.
The Hopfield network always stabilizes to a fixed point. There is a very important detail regarding the
Hopfield network to achieve this stability. In the examples thus far, we have not had a problem getting a
stable output from the network, so we have not presented this detail of network operation. This detail is the
need to update the network asynchronously. This means that changes do not occur simultaneously to outputs
that are fed back as inputs, but rather occur for one vector component at a time. The true operation of the
Hopfield network follows the procedure below for input vector Invec and output vector Outvec:
1. Apply an input, Invec, to the network, and initialize Outvec = Invec

2. Start with i = 1
3. Calculate Value i = DotProduct ( Inveci, Columni of Weight matrix)
4. Calculate Outvec i = f(Valuei) where f is the threshold function discussed previously
5. Update the input to the network with component Outvec i
6. Increment i, and repeat steps 3, 4, 5, and 6 until Invec = Outvec (note that when i reaches its
maximum value, it is then next reset to 1 for the cycle to continue)
Now let’s see how to apply this procedure. Building on the last example, we now input E = (1, 0, 0, 1), which
is at an equal distance from A and B. Without applying the asynchronous procedure above, but instead using
the shortcut procedure we’ve been using so far, you would get an output F = (0, 1, 1, 0). This vector, F, as
subsequent input would result in E as the output. This is incorrect since the network oscillates between two
states. We have updated the entire input vector synchronously.
Now let’s apply asynchronous update. For input E, (1,0,0,1) we arrive at the following results detailed for
each update step, in Table 1.1.
Table 1.1 Example of Asynchronous Update for the Hopfield Network

StepiInvecColumn of Weight vectorValueOutvecnotes
010011001initialization : set Outvec = Invec = Input pattern
1110010 −3 3 −3−30001column 1 of Outvec changed to 0
220001−3 0 −3 330101column 2 of Outvec changed to 1
3301013 −3 0 −3−60101column 3 of Outvec stays as 0
440101−3 3 −3 030101column 4 of Outvec stays as 1
5101010 −3 3 −3−60101column 1 stable as 0
Asynchronous Update 25
620101−3 0 −3 330101column 2 stable as 1

7301013 −3 0 −3−60101column 3 stable as 0
840101−3 3 −3 030101column 4 stable as 1; stable recalled pattern = 0101
Asynchronous Update 26

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Binary and Bipolar Inputs
Two types of inputs that are used in neural networks are binary and bipolar inputs. We have already seen
examples of binary input. Bipolar inputs have one of two values, 1 and –1. There is clearly a one−to−one
mapping or correspondence between them, namely having −1 of bipolar correspond to a 0 of binary. In
determining the weight matrix in some situations where binary strings are the inputs, this mapping is used,
and when the output appears in bipolar values, the inverse transformation is applied to get the corresponding
binary string. A simple example would be that the binary string 1 0 0 1 is mapped onto the bipolar string 1 –1
–1 1; while using the inverse transformation on the bipolar string –1 1 –1 –1, we get the binary string 0 1 0 0.
Bias
The use of threshold value can take two forms. One we showed in the example. The activation is compared to
the threshold value, and the neuron fires if the threshold value is attained or exceeded. The other way is to add
a value to the activation itself, in which case it is called the bias, and then determining the output of the
neuron. We will encounter bias and gain later.
Another Example for the Hopfield Network

You will see in Chapter 12 an application of Kohonen’s feature map for pattern recognition. Here we give an
example of pattern association using a Hopfield network. The patterns are some characters. A pattern
representing a character becomes an input to a Hopfield network through a bipolar vector. This bipolar vector
is generated from the pixel (picture element) grid for the character, with an assignment of a 1 to a black pixel
and a −1 to a pixel that is white. A grid size such as 5x7 or higher is usually employed in these approaches.
The number of pixels involved will then be 35 or more, which determines the dimension of a bipolar vector
for the character pattern.
We will use, for simplicity, a 3x3 grid for character patterns in our example. This means the Hopfield network
has 9 neurons in the only layer in the network. Again for simplicity, we use two exemplar patterns, or
reference patterns, which are given in Figure 1.5. Consider the pattern on the left as a representation of the
character “plus”, +, and the one on the right that of “minus”, − .
Figure 1.5 The “plus” pattern and “minus” pattern.
The bipolar vectors that represent the characters in the figure, reading the character pixel patterns row by row,
left to right, and top to bottom, with a 1 for black and −1 for white pixels, are C+ = (−1, 1, −1, 1, 1, 1, −1, 1,
−1), and C− = (−1, −1, −1, 1, 1, 1, −1, −1, −1). The weight matrix W is:
0 0 2 −2 −2 −2 2 0 2
0 0 0 0 0 0 0 2 0
Binary and Bipolar Inputs 27

2 0 0 −2 −2 −2 2 0 2
2 0 −2 0 2 2 −2 0 −2
W= 2 0 −2 2 0 2 −2 0 −2
2 0 −2 2 2 0 −2 0 −2
2 0 2 −2 −2 −2 0 0 2
0 2 0 0 0 0 0 0 0
2 0 2 −2 −2 −2 2 0 0
The activations with input C+ are given by the vector (−12, 2, −12, 12, 12, 12, −12, 2, −12). With input C−,
the activations vector is (−12, −2, −12, 12, 12, 12, −12, −2, −12).
When this Hopfield network uses the threshold function
1 if x >= 0
f(x) = { −1 if x [le] 0
the corresponding outputs will be C+ and C−, respectively, showing the stable recall of the exemplar vectors,
and establishing an autoassociation for them. When the output vectors are used to construct the corresponding
characters, you get the original character patterns.
Let us now input the character pattern in Figure 1.6.
Figure 1.6 Corrupted “minus” pattern.
We will call the corresponding bipolar vector A = (1, −1, −1, 1, 1, 1, −1, −1, −1). You get the activation
vector (−12, −2, −8, 4, 4, 4, −8, −2, −8) giving the output vector, C− = (−1, −1, −1, 1, 1, 1, −1, −1, −1). In
other words, the character −, corrupted slightly, is recalled as the character − by the Hopfield network. The
intended pattern is recognized.
We now input a bipolar vector that is different from the vectors corresponding to the exemplars, and see
whether the network can store the corresponding pattern. The vector we choose is B = (1, −1, 1, −1, −1, −1, 1,
−1, 1). The corresponding neuron activations are given by the vector (12, −2, 12, −4, −4, −4, 12, −2, 12)
which causes the output to be the vector (1, −1, 1, −1, −1, −1, 1, −1, 1), same as B. An additional pattern,
which is a 3x3 grid with only the corner pixels black, as shown in Figure 1.7, is also recalled since it is
autoassociated, by this Hopfield network.
Figure 1.7 Pattern result.
If we omit part of the pattern in Figure 1.7, leaving only the top corners black, as in Figure 1.8, we get the
bipolar vector D = (1, −1, 1, −1, −1, −1, −1, −1, −1). You can consider this also as an incomplete or corrupted
version of the pattern in Figure 1.7. The network activations turn out to be (4, −2, 4, −4, −4, −4, 8, −2, 8) and
Binary and Bipolar Inputs 28

give the output (1, −1, 1, −1, −1, −1, 1, −1, 1), which is B.
Figure 1.8 A partly lost Pattern of Figure 1.7.
Summary
In this chapter we introduced a neural network as a collection of processing elements distributed over a finite
number of layers and interconnected with positive or negative weights, depending on whether cooperation or
competition (or inhibition) is intended. The activation of a neuron is basically a weighted sum of its inputs. A
threshold function determines the output of the network. There may be layers of neurons in between the input
layer and the output layer, and some such middle layers are referred to as hidden layers, others by names such
as Grossberg or Kohonen layers, named after the researchers Stephen Grossberg and Teuvo Kohonen, who
proposed them and their function. Modification of the weights is the process of training the network, and a
network subject to this process is said to be learning during that phase of the operation of the network. In
some network operations, a feedback operation is used in which the current output is treated as modified input
to the same network.
You have seen a couple of examples of a Hopfield network, one of them for pattern recognition.
Neural networks can be used for problems that can’t be solved with a known formula and for problems with
incomplete or noisy data. Neural networks seem to have the capacity to recognize patterns in the data
presented to it, and are thus useful in many types of pattern recognition problems.
Summary 29

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Chapter 2
C++ and Object Orientation
Introduction to C++
C++ is an object−oriented programming language built on the base of the C language. This chapter gives you
a very brief introduction to C++, touching on many important aspects of C++, so you would be able to follow
our presentations of the C++ implementations of neural network models and write your own C++ programs.
The C++ language is a superset of the C language. You could write C++ programs like C programs (a few of
the programs in this book are like that), or you could take advantage of the object−oriented features of C++ to
write object−oriented programs (like the backpropagation simulator of Chapter 7). What makes a
programming language or programming methodology object oriented? Well, there are several indisputable
pillars of object orientation. These features stand out more than any other as far as object orientation goes.
They are encapsulation, data hiding, overloading, polymorphism, and the grand−daddy of them all:
inheritance. Each of the pillars of object−orientation will be discussed in the coming sections, but before we
tackle these, we need to answer the question, What does all this object−oriented stuff buy me ? By using the
object−oriented features of C++, in conjunction with Object−Oriented Analysis and Design(OOAD), which is
a methodology that fully utilizes object orientation, you can have well−packaged, reusable, extensible, and
reliable programs and program segments. It’s beyond the scope of this book to discuss OOAD, but it’s
recommended you read Booch or Rumbaugh to get more details on OOAD and how and why to change your
programming style forever! See the reference section in the back of this book for more information on these
readings. Now let’s get back to discussing the great object−oriented features of C++.
Encapsulation
In C++ you have the facility to encapsulate data and the operations that manipulate that data, in an appropriate
object. This enables the use of these collections of data and function, called objects , in programs other than
the program for which they were originally created. With objects, just as with the traditional concept of
subroutines, you make functional blocks of code. You still have language−supported abstractions such as
scope and separate compilation available. This is a rudimentary form of encapsulation. Objects carry
encapsulation a step further. With objects, you define not only the way a function operates, or its
implementation, but also the way an object can be accessed, or its interface. You can specify access
differently for different entities. For example, you could make function do_operation() contained inside
Object A accessible to Object B but not to Object C. This access qualification can also be used for data
members inside an object. The encapsulation of data and the intended operations on them prevents the data
from being subjected to operations not meant for them. This is what really makes objects reusable and
portable! The operations are usually given in the form of functions operating upon the data items. Such
functions are also called methods in some object−oriented programming languages. The data items and the
functions that manipulate them are combined into a structure called a class. A class is an abstract data type.
When you make an instance of a class, you have an object. This is no different than when you instantiate an
Chapter 2 C++ and Object Orientation 30

integer type to create variables i and j. For example, you design a class called ElectronicBook, with a data
element called ArrayofPages. When you instantiate your class you make objects of type ElectronicBook.
Suppose that you create two of these called EB_Geography and EB_History. Every object that is
instantiated has its own data member inside it, referred to by ArrayOfPages.
Data Hiding
Related to the idea of encapsulation is the concept of data hiding. Encapsulation hides the data from other
classes and functions in other classes. Going back to the ElectronicBook class, you could define functions
like GetNextPage, GetPreviousPage, and GetCurrentPage as the only means of accessing information in
the ArrayofPages data member, by functions that access the ElectronicBook object. Although there may be a
hundred and one other attributes and data elements in the class ElectronicBook, these are all hidden from
view. This makes programs more reliable, since publishing a specific interface to an object prevents
inadvertent access to data in ways that were not designed or accounted for. In C++, the access to an object,
and its encapsulated data and functions is treated very carefully, by the use of keywords private, protected,
and public. One has the opportunity to make access specifications for data objects and functions as being
private, or protected, or public while defining a class. Only when the declaration is made as public do other
functions and objects have access to the object and its components without question. On the other hand, if the
declaration happens to be as private, there is no possibility of such access. When the declaration given is as
protected, then the access to data and functions in a class by others is not as free as when it is public, nor as
restricted as when it is private. You can declare one class as derived from another class, which will be
discussed shortly. So−called derived classes and the declaring class do get the access to the components of the
object that are declared protected. One class that is not a derived class of a second class can get access to data
items and functions of the second class if it is declared as a friend class in the second. The three types of
declarations of access specification can be different for different components of an object. For example, some
of the data items could be declared public, some private, and the others protected. The same situation can
occur with the functions in an object. When no explicit declaration is made, the default specification is as
private.
Data Hiding 31

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Constructors and Destructors as Special Functions of C++

Constructors and destructors are special functions in C++. They define how an object is created and
destroyed. You cannot have a class defined in a C++ program without declaring and defining at least one
constructor for it. You may omit declaring and then defining a destructor only because the compiler you use
will create a default destructor. More than one constructor, but only one destructor, can be declared for a class.
Constructors are for the creation of an object of a class and for initializing it. C++ requires that every function
has a return type. The only exceptions are constructors and destructors. A constructor is given the same name
as the class for which it is a constructor. It may take arguments or it may not need them. Different constructors
for the same class differ in the number and types of arguments they take. It is a good idea to provide for each
class at least a default constructor that does not take any arguments and does not do anything except create an
object of that class type. A constructor is called at the time an object of its class is needed to be created.
A destructor also is given the same name as the class for which it is a destructor, but with the tilde (~)
preceding the name. Typically, what is done in a destructor is to have statements that ask the system to delete
the various data structures created for the class. This helps to free−up allocated memory for those data
structures. A destructor is called when the object created earlier is no longer needed in the program.
Dynamic Memory Allocation

C++ has keywords new and delete, which are used as a pair in that order, though separated by other statements
of the program. They are for making dynamic allocation of memory at the time of creation of a class object
and for freeing−up such allocated memory when it is no longer needed. You create space on the heap with the
use of new. This obviates the need in C++ for malloc, which is the function for dynamic memory allocation
used in C.
Overloading
Encapsulation of data and functions would also allow you to use the same function name in two different
objects. The use of a name for a function more than once does not have to be only in different object
declarations. Within the same object one can use the same name for functions with different functionality, if
they can be distinguished in terms of either their return type or in terms of their argument types and number.
This feature is called overloading. For example, if two different types of variables are data items in an object,
a commonly named function can be the addition, one for each of the two types of variables—thus taking
advantage of overloading. Then the function addition is said to be overloaded. But remember that the function
main is just about the only function that cannot be overloaded.
Polymorphism and Polymorphic Functions

A polymorphic function is a function whose name is used in different ways in a program. It can be also
Constructors and Destructors as Special Functions of C++ 32

declared virtual, if the intention is late binding. This enables it to be bound at run time. Late binding is also
referred to as dynamic binding. An advantage in declaring a function in an object as virtual is that, if the
program that uses this object calls that function only conditionally, there is no need to bind the function early,
during the compilation of the program. It will be bound only if the condition is met and the call of the function
takes place. For example, you could have a polymorphic function called draw() that is associated with
different graphical objects, like for rectangle, circle, and sphere. The details or methods of the functions are
different, but the name draw() is common. If you now have a collection of these objects and pick up an
arbitrary object without knowing exactly what it is (via a pointer, for example), you can still invoke the draw
function for the object and be assured that the right draw function will be bound to the object and called.
Overloading Operators
You can overload operators in addition to overloading functions. As a matter of fact, the system defined left
shift operator << is also overloaded in C++ when used with cout, the C++ variation of the C language printf
function. There is a similar situation with the right shift operator >> in C++ when used with cin, the C++
variation of the C language scanf function. You can take any operator and overload it. But you want to be
cautious and not overdo it, and also you do not create confusion when you overload an operator. The guiding
principle in this regard is the creation of a code−saving and time−saving facility while maintaining simplicity
and clarity. Operator overloading is especially useful for doing normal arithmetic on nonstandard data types.
You could overload the multiplication symbol to work with complex numbers, for example.
Overloading Operators 33

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Inheritance
The primary distinction for C++ from C is that C++ has classes. Objects are defined in classes. Classes
themselves can be data items in other classes, in which case one class would be an element of another class.
Of course, then one class is a member, which brings with it its own data and functions, in the second class.
This type of relationship is referred to as a “has−a” relationship: Object A has an Object B inside it.
A relationship between classes can be established not only by making one class a member of another but also
by the process of deriving one class from another. One class can be derived from another class, which
becomes its base class. Then a hierarchy of classes is established, and a sort of parent–child relationship
between classes is established. The derived class inherits, from the base class, some of the data members and
functions. This type of relationship is referred to as an “is−a” relationship. You could have class Rectangle be
derived from class Shape, since Rectangle is a Shape. Naturally, if a class A is derived from a class B, and if
B itself is derived from a class C, then A inherits from both B and C. A class can be derived from more than
one class. This is how multiple inheritance occurs. Inheritance is a powerful mechanism for creating base
functionality that is passed onto next generations for further enhancement or modification.
Derived Classes
When one class has some members declared in it as protected, then such members would be hidden from
other classes, but not from the derived classes. In other words, deriving one class from another is a way of
accessing the protected members of the parent class by the derived class. We then say that the derived class is
inheriting from the parent class those members in the parent class that are declared as protected or public.
In declaring a derived class from another class, access or visibility specification can be made, meaning that
such derivation can be public or the default case, private. Table 2.1 shows the consequences of such
specification when deriving one class from another.
Table 2.1 Visibility of Base Class Members in Derived Class

Derivation SpecificationBase Class SpecificationDerived Class Access
private(default)privatenone
protectedfull access, private in derived class
publicfull access, public in derived class
publicprivatenone
protectedfull access, protected in derived class
publicfull access, public in derived class
Reuse of Code
C++ is also attractive for the extendibility of the programs written in it and for the reuse opportunity, thanks to
the features in C++ such as inheritance and polymorphism mentioned earlier. A new programming project
cannot only reuse classes that are created for some other program, if they are appropriate, but can extend
Inheritance 34
another program with additional classes and functions as deemed necessary. You could inherit from an
existing class hierarchy and only change functionality where you need to.
C++ Compilers
All of the programs in this book have been compiled and tested with Turbo C++, Borland C++, Microsoft
C/C++, and Microsoft Visual C++. These are a few of the popular commercial C++ compilers available. You
should be able to use most other commercial C++ compilers also. All of the programs should also port easily
to other operating systems like Unix and the Mac, because they are all character−based programs.
C++ Compilers 35

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Writing C++ Programs

Before one starts writing a C++ program for a particular problem, one has to have a clear picture of the
various parameters and variables that would be part of the problem definition and/or its solution. In addition,
it should be clear as to what manipulations need to be performed during the solution process. Then one
carefully determines what classes are needed and what relationships they have to each other in a hierarchy of
classes. Think about is−a and has−a relationships to see where classes need to defined, and which classes
could be derived from others. It would be far more clear to the programmer at this point in the program plan
what the data and function access specifications should be and so on. The typical compilation error messages
a programmer to C++ may encounter are stating that a particular function or data is not a member of a
particular class, or that it is not accessible to a class, or that a constructor was not available for a particular
class. When function arguments at declaration and at the place the function is called do not match, either for
number or for types or both, the compiler thinks of them as two different functions. The compiler does not
find the definition and/or declaration of one of the two and has reason to complain. This type of error in one
line of code may cause the compiler to alert you that several other errors are also present, perhaps some in
terms of improper punctuation. In that case, remedying the fundamental error that was pointed out would
straighten many of the other argued matters.
The following list contains a few additional particulars you need to keep in mind when writing C++ programs.
• A member x of an object A is referred to with A.x just as done with structure elements in C.
• If you declare a class B, then the constructor function is also named B. B has no return type. If
this constructor takes, say, one argument of type integer, you define the constructor using the syntax:
B::B(int){whatever the function does};
• If you declare a member function C of class B, where return type of C is, say, float and C takes
two arguments, one of type float, and the other int, then you define C with the syntax: float
B::C(float,int){whatever the function does};
• If you declare a member function D of class B, where D does not return any value and takes no
arguments, you define D using the syntax: void B::D( ){whatever the function does};
• If G is a class derived from, say, class B previously mentioned, you declare G using the syntax:
class G:B. The constructor for G is defined using the syntax: G::G(arguments of
G):B(int){whatever the function does}. If, on the other hand, G is derived from, say, class B as well
as class T, then you declare G using the syntax: class G:B,T.
• If one class is declared as derived from more than one other class, that is, if there are more than
one base class for it, the derivations specification can be different or the same. Thus the class may be
derived from one class publicly and at the same time from another class privately.
• If you have declared a global variable y external to a class B, and if you also have a data member
y in the class B, you can use the external y with the reference symbol ::. Thus ::y refers to the global
variable, whereas y, within a member function of B, or B.y refers to the data member of B. This way
polymorphic functions can also be distinguished from each other.
Writing C++ Programs 36

This is by no means a comprehensive list of features, but more a brief list of important constructs in C++. You
will see examples of C++ usage in later chapters.
Summary
A few highlights of the C++ language are presented.
• C++ is an object−oriented language with full compatibility with the C language.

• You create classes in C++ that encapsulate data and functions that operate on the data, and hiding
data where a public interface is not needed.
• You can create hierarchies of classes with the facility of inheritance. Polymorphism is a feature
that allows you to apply a function to a task according to the object the function is operating on.
• Another feature in C++ is overloading of operators, which allows you to create new functionality
for existing operators in a different context.
• Overall, C++ is a powerful language fitting the object−oriented paradigm that enables software
reuse and enhanced reliability.
Summary 37

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Chapter 3
A Look at Fuzzy Logic
Crisp or Fuzzy Logic?
Logic deals with true and false. A proposition can be true on one occasion and false on another. “Apple is a
red fruit” is such a proposition. If you are holding a Granny Smith apple that is green, the proposition that
apple is a red fruit is false. On the other hand, if your apple is of a red delicious variety, it is a red fruit and the
proposition in reference is true. If a proposition is true, it has a truth value of 1; if it is false, its truth value is
0. These are the only possible truth values. Propositions can be combined to generate other propositions, by
means of logical operations.
When you say it will rain today or that you will have an outdoor picnic today, you are making statements with
certainty. Of course your statements in this case can be either true or false. The truth values of your statements
can be only 1, or 0. Your statements then can be said to be crisp.
On the other hand, there are statements you cannot make with such certainty. You may be saying that you
think it will rain today. If pressed further, you may be able to say with a degree of certainty in your statement
that it will rain today. Your level of certainty, however, is about 0.8, rather than 1. This type of situation is
what fuzzy logic was developed to model. Fuzzy logic deals with propositions that can be true to a certain
degree—somewhere from 0 to 1. Therefore, a proposition’s truth value indicates the degree of certainty about
which the proposition is true. The degree of certainity sounds like a probability (perhaps subjective
probability), but it is not quite the same. Probabilities for mutually exclusive events cannot add up to more
than 1, but their fuzzy values may. Suppose that the probability of a cup of coffee being hot is 0.8 and the
probability of the cup of coffee being cold is 0.2. These probabilities must add up to 1.0. Fuzzy values do not
need to add up to 1.0. The truth value of a proposition that a cup of coffee is hot is 0.8. The truth value of a
proposition that the cup of coffee is cold can be 0.5. There is no restriction on what these truth values must
add up to.
Fuzzy Sets
Fuzzy logic is best understood in the context of set membership. Suppose you are assembling a set of rainy
days. Would you put today in the set? When you deal only with crisp statements that are either true or false,
your inclusion of today in the set of rainy days is based on certainty. When dealing with fuzzy logic, you
would include today in the set of rainy days via an ordered pair, such as (today, 0.8). The first member in
such an ordered pair is a candidate for inclusion in the set, and the second member is a value between 0 and 1,
inclusive, called the degree of membership in the set. The inclusion of the degree of membership in the set
makes it convenient for developers to come up with a set theory based on fuzzy logic, just as regular set
theory is developed. Fuzzy sets are sets in which members are presented as ordered pairs that include
information on degree of membership. A traditional set of, say, k elements, is a special case of a fuzzy set,
where each of those k elements has 1 for the degree of membership, and every other element in the universal
Chapter 3 A Look at Fuzzy Logic 38

set has a degree of membership 0, for which reason you don’t bother to list it.
Fuzzy Set Operations

The usual operations you can perform on ordinary sets are union, in which you take all the elements that are in
one set or the other; and intersection, in which you take the elements that are in both sets. In the case of fuzzy
sets, taking a union is finding the degree of membership that an element should have in the new fuzzy set,
which is the union of two fuzzy sets.
If a, b, c, and d are such that their degrees of membership in the fuzzy set A are 0.9, 0.4, 0.5, and 0,
respectively, then the fuzzy set A is given by the fit vector (0.9, 0.4, 0.5, 0). The components of this fit vector
are called fit values of a, b, c, and d.
Union of Fuzzy Sets
Consider a union of two traditional sets and an element that belongs to only one of those sets. Earlier you saw
that if you treat these sets as fuzzy sets, this element has a degree of membership of 1 in one case and 0 in the
other since it belongs to one set and not the other. Yet you are going to put this element in the union. The
criterion you use in this action has to do with degrees of membership. You need to look at the two degrees of
membership, namely, 0 and 1, and pick the higher value of the two, namely, 1. In other words, what you want
for the degree of membership of an element when listed in the union of two fuzzy sets, is the maximum value
of its degrees of membership within the two fuzzy sets forming a union.
If a, b, c, and d have the respective degrees of membership in fuzzy sets A, B as A = (0.9, 0.4, 0.5, 0) and B =
(0.7, 0.6, 0.3, 0.8), then A [cup] B = (0.9, 0.6, 0.5, 0.8).
Fuzzy Set Operations 39


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Intersection and Complement of Two Fuzzy Sets
Analogously, the degree of membership of an element in the intersection of two fuzzy sets is the minimum, or
the smaller value of its degree of membership individually in the two sets forming the intersection. For
example, if today has 0.8 for degree of membership in the set of rainy days and 0.5 for degree of membership
in the set of days of work completion, then today belongs to the set of rainy days on which work is completed
to a degree of 0.5, the smaller of 0.5 and 0.8.
Recall the fuzzy sets A and B in the previous example. A = (0.9, 0.4, 0.5, 0) and B = (0.7, 0.6, 0.3, 0.8).
A[cap]B, which is the intersection of the fuzzy sets A and B, is obtained by taking, in each component, the
smaller of the values found in that component in A and in B. Thus A[cap]B = (0.7, 0.4, 0.3, 0).
The idea of a universal set is implicit in dealing with traditional sets. For example, if you talk of the set of
married persons, the universal set is the set of all persons. Every other set you consider in that context is a
subset of the universal set. We bring up this matter of universal set because when you make the complement
of a traditional set A, you need to put in every element in the universal set that is not in A. The complement of
a fuzzy set, however, is obtained as follows. In the case of fuzzy sets, if the degree of membership is 0.8 for a
member, then that member is not in that set to a degree of 1.0 – 0.8 = 0.2. So you can set the degree of
membership in the complement fuzzy set to the complement with respect to 1. If we return to the scenario of
having a degree of 0.8 in the set of rainy days, then today has to have 0.2 membership degree in the set of
nonrainy or clear days.
Continuing with our example of fuzzy sets A and B, and denoting the complement of A by A’, we have A’ =
(0.1, 0.6, 0.5, 1) and B’ = (0.3, 0.4, 0.7, 0.2). Note that A’ [cup] B’ = (0.3, 0.6, 0.7, 1), which is also the
complement of A [cap] B. You can similarly verify that the complement of A [cup] B is the same as A’ [cap]
B’. Furthermore, A [cup] A’ = (0.9, 0.6, 0.5, 1) and A [cap] A’ = (0.1, 0.4, 0.5, 0), which is not a vector of
zeros only, as would be the case in conventional sets. In fact, A and A’ will be equal in the sense that their fit
vectors are the same, if each component in the fit vector is equal to 0.5.

Applications of fuzzy sets and fuzzy logic are found in many fields, including artificial intelligence,
engineering, computer science, operations research, robotics, and pattern recognition. These fields are also
ripe for applications for neural networks. So it seems natural that fuzziness should be introduced in neural
networks themselves. Any area where humans need to indulge in making decisions, fuzzy sets can find a
place, since information on which decisions are to be based may not always be complete and the reliability of
the supposed values of the underlying parameters is not always certain.
Examples of Fuzzy Logic
Let us say five tasks have to be performed in a given period of time, and each task requires one person
dedicated to it. Suppose there are six people capable of doing these tasks. As you have more than enough
Intersection and Complement of Two Fuzzy Sets 40

people, there is no problem in scheduling this work and getting it done. Of course who gets assigned to which
task depends on some criterion, such as total time for completion, on which some optimization can be done.
But suppose these six people are not necessarily available during the particular period of time in question.
Suddenly, the equation is seen in less than crisp terms. The availability of the people is fuzzy−valued. Here is
an example of an assignment problem where fuzzy sets can be used.
Commercial Applications
Many commercial uses of fuzzy logic exist today. A few examples are listed here:
• A subway in Sendai, Japan uses a fuzzy controller to control a subway car. This controller has
outperformed human and conventional controllers in giving a smooth ride to passengers in all terrain
and external conditions.
• Cameras and camcorders use fuzzy logic to adjust autofocus mechanisms and to cancel the jitter
caused by a shaking hand.
• Some automobiles use fuzzy logic for different control applications. Nissan has patents on fuzzy
logic braking systems, transmission controls, and fuel injectors. GM uses a fuzzy transmission system
in its Saturn vehicles.
• FuziWare has developed and patented a fuzzy spreadsheet called FuziCalc that allows users to
incorporate fuzziness in their data.
• Software applications to search and match images for certain pixel regions of interest have been
developed. Avian Systems has a software package called FullPixelSearch.
• A stock market charting and research tool called SuperCharts from Omega Research, uses fuzzy
logic in one of its modules to determine whether the market is bullish, bearish, or neutral.
Commercial Applications 41

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95

There are a number of ways fuzzy logic can be used with neural networks. Perhaps the simplest way is to use
a fuzzifier function to preprocess or post−process data for a neural network. This is shown in Figure 3.1,
where a neural network has a preprocessing fuzzifier that converts data into fuzzy data for application to a
neural network.
Figure 3.1 A neural network with fuzzy preprocessor.
Let us build a simple fuzzifier based on an application to predict the direction of the stock market. Suppose
that you wish to fuzzify one set of data used in the network, the Federal Reserve’s fiscal policy, in one of four
fuzzy categories: very accommodative, accommodative, tight, or very tight. Let us suppose that the raw data
that we need to fuzzify is the discount rate and the interest rate that the Federal Reserve controls to set the
fiscal policy. Now, a low discount rate usually indicates a loose fiscal policy, but this depends not only on the
observer, but also on the political climate. There is a probability, for a given discount rate that you will find
two people who offer different categories for the Fed fiscal policy. Hence, it is appropriate to fuzzify the data,
so that the data we present to the neural network is like what an observer would see.
Figure 3.2 shows the fuzzy categories for different interest rates. Note that the category tight has the largest
range. At any given interest rate level, you could have one possible category or several. If only one interest
rate is present on the graph, this indicates that membership in that fuzzy set is 1.0. If you have three possible
fuzzy sets, there is a requirement that membership add up to 1.0. For an interest rate of 8%, you have some
chance of finding this in the tight category or the accommodative category. To find out the percentage
probability from the graph, take the height of each curve at a given interest rate and normalize this to a
one−unit length. At 8%, the tight category is about 0.8 unit in height, and accommodative is about 0.3 unit in
height. The total is about 1.1 units, and the probability of the value being tight is then 0.8/1.1 = 0.73, while the
probability of the value being accommodative is 0.27.
Figure 3.2 Fuzzy categories for Federal Reserve policy based on the Fed discount rate.
Fuzziness in Neural Networks 42

Code for the Fuzzifier

Let’s develop C++ code to create a simple fuzzifier. A class called category is defined in Listing 3.1. This
class encapsulates the data that we need to define, the categories in Figure 3.2. There are three private data
members called lowval, midval, and highval. These represent the values on the graph that define the category
triangle. In the tight category, the lowval is 5.0, the midval is 8.5, and the highval is 12.0. The category class
allows you to instantiate a category object and assign parameters to it to define it. Also, there is a string called
name that identifies the category, e.g. “tight.” Various member functions are used to interface to the private
data members. There is setval(), for example, which lets you set the value of the three parameters, while
gethighval() returns the value of the parameter highval. The function getshare() returns the relative value of
membership in a category given an input. In the example discussed earlier, with the number 8.0 as the Fed
discount rate and the category tight defined according to the graph in Figure 3.2, getshare() would return 0.8.
Note that this is not yet normalized. Following this example, the getshare() value from the accommodative
category would also be used to determine the membership weights. These weights define a probability in a
given category. A random number generator is used to define a value that is used to select a fuzzy category
based on the probabilities defined.
Listing 3.1 fuzzfier.h
// fuzzfier.h V. Rao, H. Rao

// program to fuzzify data
class category
{
private:
char name[30];
float lowval,highval,midval;
public:
category(){};
void setname(char *);
char * getname();
void setval(float&,float&,float&);
float getlowval();
float getmidval();
float gethighval();
float getshare(const float&);
~category(){};
};
int randnum(int);
Code for the Fuzzifier 43


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Let’s look at the implementation file in Listing 3.2.
Listing 3.2 fuzzfier.cpp
// fuzzfier.cpp V. Rao, H. Rao

// program to fuzzify data
#include <iostream.h>
#include <stdlib.h>
#include <time.h>
#include <string.h>
#include <fuzzfier.h>
void category::setname(char *n)

{
strcpy(name,n);
}
char * category::getname()
{
return name;
}
void category::setval(float &h, float &m, float &l)

{
highval=h;
midval=m;
lowval=l;
}
float category::getlowval()
{
return lowval;
}
float category::getmidval()
{
return midval;
}
float category::gethighval()
{
return highval;
}
float category::getshare(const float & input)

{
// this member function returns the relative membership
// of an input in a category, with a maximum of 1.0
float output;

float midlow, highmid;
midlow=midval−lowval;
highmid=highval−midval;
// if outside the range, then output=0

if ((input <= lowval) || (input >= highval))
output=0;
else
{
if (input > midval)
output=(highval−input)/highmid;
else
if (input==midval)
output=1.0;
else
output=(input−lowval)/midlow;
}
return output;
int randomnum(int maxval)

{
// random number generator
// will return an integer up to maxval
srand ((unsigned)time(NULL));
return rand() % maxval;
}
void main()
{
// a fuzzifier program that takes category information:
// lowval, midval and highval and category name
// and fuzzifies an input based on
// the total number of categories and the membership
// in each category
int i=0,j=0,numcat=0,randnum;
float l,m,h, inval=1.0;
char input[30]=" ";

category * ptr[10];
float relprob[10];
float total=0, runtotal=0;
//input the category information; terminate with `done';
while (1)

cout << "\nPlease type in a category name, e.g. Cool\n";

cout << "Enter one word without spaces\n";
cout << "When you are done, type `done' :\n\n";
ptr[i]= new category;

cin >> input;
if ((input[0]=='d' && input[1]=='o' &&

input[2]=='n' && input[3]=='e')) break;
ptr[i]−>setname(input);
cout << "\nType in the lowval, midval and highval\n";

cout << "for each category, separated by spaces\n";
cout << " e.g. 1.0 3.0 5.0 :\n\n";
cin >> l >> m >> h;

ptr[i]−>setval(h,m,l);
i++;
numcat=i; // number of categories
// Categories set up: Now input the data to fuzzify

cout <<"\n\n";
cout << "===================================\n";
cout << "==Fuzzifier is ready for data==\n";
cout << "===================================\n";
while (1)
{
cout << "\ninput a data value, type 0 to terminate\n";
cin >> inval;
if (inval == 0) break;
// calculate relative probabilities of

// input being in each category
total=0;
for (j=0;j<numcat;j++)
{
relprob[j]=100*ptr[j]−>getshare(inval);
total+=relprob[j];
}
if (total==0)
{
cout << "data out of range\n";
exit(1);
}
randnum=randomnum((int)total);
j=0;
runtotal=relprob[0];
while ((runtotal<randnum)&&(j<numcat))

{
j++;
runtotal += relprob[j];
}
cout << "\nOutput fuzzy category is ==> " <<

ptr[j]−>getname()<<"<== \n";
cout <<"category\t"<<"membership\n";
cout <<"−−−−−−−−−−−−−−−\n";
for (j=0;j<numcat;j++)
{
cout << ptr[j]−>getname()<<"\t\t"<<
(relprob[j]/total) <<"\n";
}
cout << "\n\nAll done. Have a fuzzy day !\n";
This program first sets up all the categories you define. These could be for the example we choose or any
example you can think of. After the categories are defined, you can start entering data to be fuzzified. As you
enter data you see the probability aspect come into play. If you enter the same value twice, you may end up
with different categories! You will see sample output shortly, but first a technical note on how the weighted
probabilities are set up. The best way to explain it is with an example. Suppose that you have defined three
categories, A, B, and C. Suppose that category A has a relative membership of 0.8, category B of 0.4, and
category C of 0.2. In the program, these numbers are first multiplied by 100, so you end up with A=80, B=40,
and C=20. Now these are stored in a vector with an index j initialized to point to the first category. Let’s say
that these three numbers represent three adjacent number bins that are joined together. Now pick a random
number to index into the bin that has its maximum value of (80+40+20). If the number is 100, then it is
greater than 80 and less than (80+40), you end up in the second bin that represents B. Does this scheme give
you weighted probabilities? Yes it does, since the size of the bin (given a uniform distribution of random
indexes into it) determines the probability of falling into the bin. Therefore, the probability of falling into bin
A is 80/(80+40+20).


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Sample output from the program is shown below. Our input is in italic; computer output is not. The categories
defined by the graph in Figure 3.2 are entered in this example. Once the categories are set up, the first data
entry of 4.0 gets fuzzified to the accommodative category. Note that the memberships are also presented in
each category. The same value is entered again, and this time it gets fuzzified to the very accommodative
category. For the last data entry of 12.5, you see that only the very tight category holds membership for this
value. In all cases you will note that the memberships add up to 1.0.
fuzzfier
Please type in a category name, e.g. Cool

Enter one word without spaces
When you are done, type `done' :
v.accommodative
Type in the lowval, midval and highval

for each category, separated by spaces
e.g. 1.0 3.0 5.0 :
0 3 6

accommodative

e.g. 1.0 3.0 5.0 :
3 6 9

tight

e.g. 1.0 3.0 5.0 :
5 8.5 12

v.tight


e.g. 1.0 3.0 5.0 :
10 12 14

done
===================================
==Fuzzifier is ready for data==
===================================
input a data value, type 0 to terminate

4.0
Output fuzzy category is ==> accommodative<==

category membership
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
v.accommodative 0.666667
accommodative 0.333333
tight 0
v.tight 0

4.0
Output fuzzy category is ==> v.accommodative<==

category membership
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
v.accommodative 0.666667
tight 0
v.tight 0

7.5
Output fuzzy category is ==> accommodative<==

category membership
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
v.accommodative 0
tight 0.588235
v.tight 0

11.0
Output fuzzy category is ==> tight<==

category membership
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
v.accommodative 0
accommodative 0
tight 0.363636
v.tight 0.636364

12.5
Output fuzzy category is ==> v.tight<==

category membership
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
v.accommodative 0
accommodative 0
tight 0
v.tight 1

0
All done. Have a fuzzy day !
Fuzzy Control Systems

The most widespread use of fuzzy logic today is in fuzzy control applications. You can use fuzzy logic to
make your air conditioner cool your room. Or you can design a subway system to use fuzzy logic to control
the braking system for smooth and accurate stops. A control system is a closed−loop system that typically
controls a machine to achieve a particular desired response, given a number of environmental inputs. A fuzzy
control system is a closed−loop system that uses the process of fuzzification, as shown in the Federal Reserve
policy program example, to generate fuzzy inputs to an inference engine, which is a knowledge base of
actions to take. The inverse process, called defuzzification, is also used in a fuzzy control system to create
crisp, real values to apply to the machine or process under control. In Japan, fuzzy controllers have been used
to control many machines, including washing machines and camcorders.
Figure 3.3 shows a diagram of a fuzzy control system. The major parts of this closed−loop system are:
Figure 3.3 Diagram of a fuzzy control system.
• machine under control—this is the machine or process that you are controlling, for example, a
washing machine
• outputs—these are the measured response behaviors of your machine, for example, the
temperature of the water
• fuzzy outputs—these are the same outputs passed through a fuzzifier, for example, hot or very
cold
• inference engine/fuzzy rule base—an inference engine converts fuzzy outputs to actions to take
by accessing fuzzy rules in a fuzzy rule base. An example of a fuzzy rule: IF the output is very cold,
THEN increase the water temperature setting by a very large amount
• fuzzy inputs—these are the fuzzy actions to perform, such as increase the water temperature
setting by a very large amount
• inputs—these are the (crisp) dials on the machine to control its behavior, for example, water
temperature setting = 3.423, converted from fuzzy inputs with a defuzzifier
The key to development of a fuzzy control system is to iteratively construct a fuzzy rule base that yields the
desired response from your machine. You construct these fuzzy rules from knowledge about the problem. In
many cases this is very intuitive and gives you a robust control system in a very short amount of time.
Fuzzy Control Systems 50

Fuzzy Control Systems 51


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95

Fuzziness can enter neural networks to define the weights from fuzzy sets. A comparison between expert
systems and fuzzy systems is important to understand in the context of neural networks. Expert systems are
based on crisp rules. Such crisp rules may not always be available. Expert systems have to consider an
exhaustive set of possibilities. Such sets may not be known beforehand. When crisp rules are not possible, and
when it is not known if the possibilities are exhaustive, the expert systems approach is not a good one.
Some neural networks, through the features of training and learning, can function in the presence of
unexpected situations. Therein neural networks have an advantage over expert systems, and they can manage
with far less information than expert systems need.
One form of fuzziness in neural networks is called a fuzzy cognitive map. A fuzzy cognitive map is like a
dynamic state machine with fuzzy states. A traditional state machine is a machine with defined states and
outputs associated with each state. Transitions from state to state take place according to input events or
stimuli. A fuzzy cognitive map looks like a state machine but has fuzzy states (not just 1 or 0). You have a set
of weights along each transition path, and these weights can be learned from a set of training data.
Our treatment of fuzziness in neural networks is with the discussion of the fuzzy associative memory,
abbreviated as FAM, which, like the fuzzy cognitive map, was developed by Bart Kosko. The FAM and the
C++ implementation are discussed in Chapter 9.
Neural−Trained Fuzzy Systems

So far we have considered how fuzzy logic plays a role in neural networks. The converse relationship, neural
networks in fuzzy systems, is also an active area of research. In order to build a fuzzy system, you must have a
set of membership rules for fuzzy categories. It is sometimes difficult to deduce these membership rules with
a given set of complex data. Why not use a neural network to define the fuzzy rules for you? A neural network
is good at discovering relationships and patterns in data and can be used to preprocess data in a fuzzy system.
Further, a neural network that can learn new relationships with new input data can be used to refine fuzzy
rules to create a fuzzy adaptive system. Neural trained fuzzy systems are being used in many commercial
applications, especially in Japan:
• The Laboratory for International Fuzzy Engineering Research (LIFE) in Yokohama, Japan has a
backpropagation neural network that derives fuzzy rules and membership functions. The LIFE system
has been successfully applied to a foreign−exchange trade support system with approximately 5000
fuzzy rules.
• Ford Motor Company has developed trainable fuzzy systems for automobile idle−speed control.
• National Semiconductor Corporation has a software product called NeuFuz that supports the
generation of fuzzy rules with a neural network for control applications.
• A number of Japanese consumer and industrial products use neural networks with fuzzy systems,
including vacuum cleaners, rice cookers, washing machines, and photocopying machines.
Fuzziness in Neural Networks 52

• AEG Corporation of Germany uses a neural−network−trained fuzzy control system for its water−
and energy−conserving washing machine. After the machine is loaded with laundry, it measures the
water level with a pressure sensor and infers the amount of laundry in the machine by the speed and
volume of water. A total of 157 rules were generated by a neural network that was trained on data
correlating the amount of laundry with the measurement of water level on the sensor.
Summary
In this chapter, you read about fuzzy logic, fuzzy sets, and simple operations on fuzzy sets. Fuzzy logic,
unlike Boolean logic, has more than two on or off categories to describe behavior of systems. You use
membership values for data in fuzzy categories, which may overlap. In this chapter, you also developed a
fuzzifier program in C++ that takes crisp values and converts them to fuzzy values, based on categories and
memberships that you define. For use with neural networks, fuzzy logic can serve as a post−processing or
pre−processing filter. Kosko developed neural networks that use fuzziness and called them fuzzy associative
memories, which will be discussed in later chapters. You also read about how neural networks can be used in
fuzzy systems to define membership functions and fuzzy rules.
Summary 53

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Chapter 4
Constructing a Neural Network
First Example for C++ Implementation
The neural network we presented in Chapter 1 is an example of a Hopfield network with a single layer. Now
we present a C++ implementation of this network. Suppose we place four neurons, all connected to one
another on this layer, as shown in Figure 4.1. Some of these connections have a positive weight and the rest
have a negative weight. You may recall from the earlier presentation of this example, that we used two input
patterns to determine the weight matrix. The network recalls them when the inputs are presented to the
network, one at a time. These inputs are binary and orthogonal so that their stable recall is assured. Each
component of a binary input pattern is either a 0 or a 1. Two vectors are orthogonal when their dot
product—the sum of the products of their corresponding components—is zero. An example of a binary input
pattern is 1 0 1 0 0. An example of a pair of orthogonal vectors is (0, 1, 0, 0, 1) and (1, 0, 0, 1, 0). An example
of a pair of vectors that are not orthogonal is (0, 1, 0, 0, 1) and (1, 1, 0, 1, 0). These last two vectors have a dot
product of 1, different from 0.
Figure 4.1 Layout of a Hopfield Network
The two patterns we want the network to have stable recall for are A = (1, 0, 1, 0) and B = (0, 1, 0, 1). The
weight matrix W is given as follows:
0 −3 3 −3
W = −3 0 −3 3
3 −3 0 −3
−3 3 −3 0
NOTE: The positive links (values with positive signs) tend to encourage agreement in a
stable configuration, whereas negative links (values with negative signs) tend to discourage
agreement in a stable configuration.
We need a threshold function also, and we define it using a threshold value, [theta], as follows:
1 if t >= [theta]
f(t) = { 0 if t < [theta]
Chapter 4 Constructing a Neural Network 54

The threshold value [theta] is used as a cut−off value for the activation of a neuron to enable it to fire. The
activation should equal or exceed the threshold value for the neuron to fire, meaning to have output 1. For our
Hopfield network, [theta] is taken as 0. There are four neurons in the only layer in this network. The first
node’s output is the output of the threshold function. The argument for the threshold function is the
activation of the node. And the activation of the node is the dot product of the input vector and the first
column of the weight matrix. So if the input vector is A, the dot product becomes 3, and f(3) = 1. And the dot
products of the second, third, and fourth nodes become –6, 3, and –6, respectively. The corresponding outputs
therefore are 0, 1, and 0. This means that the output of the network is the vector (1, 0, 1, 0), which is the same
as the input pattern. Therefore, the network has recalled the pattern as presented. When B is presented, the dot
product obtained at the first node is –6 and the output is 0. The activations of all the four nodes together with
the threshold function give (0, 1, 0, 1) as output from the network, which means that the network recalled B
as well. The weight matrix worked well with both input patterns, and we do not need to modify it.
Classes in C++ Implementation
In our C++ implementation of this network, there are the following classes: a network class, and a neuron
class. In our implementation, we create the network with four neurons, and these four neurons are all
connected to one another. A neuron is not self−connected, though. That is, there is no edge in the directed
graph representing the network, where the edge is from one node to itself. But for simplicity, we could
pretend that such a connection exists carrying a weight of 0, so that the weight matrix has 0’s in its principal
diagonal.
The functions that determine the neuron activations and the network output are declared public. Therefore
they are visible and accessible without restriction. The activations of the neurons are calculated with functions
defined in the neuron class. When there are more than one layer in a neural network, the outputs of neurons in
one layer become the inputs for neurons in the next layer. In order to facilitate passing the outputs from one
layer as inputs to another layer, our C++ implementations compute the neuron outputs in the network class.
For this reason the threshold function is made a member of the network class. We do this for the Hopfield
network as well. To see if the network has achieved correct recall, you make comparisons between the
presented pattern and the network output, component by component.
Classes in C++ Implementation 55


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
C++ Program for a Hopfield Network

For convenience every C++ program has two components: One is the header file with all of the class
declarations and lists of include library files; the other is the source file that includes the header file and the
detailed descriptions of the member functions of the classes declared in the header file. You also put the
function main in the source file. Most of the computations are done by class member functions, when class
objects are created in the function main, and calls are made to the appropriate functions. The header file has
an .h (or .hpp) extension, as you know, and the source file has a .cpp extension, to indicate that it is a C++
code file. It is possible to have the contents of the header file written at the beginning of the .cpp file and work
with one file only, but separating the declarations and implementations into two files allows you to change the
implementation of a class(.cpp) without changing the interface to the class (.h).
Header File for C++ Program for Hopfield Network
Listing 4.1 contains Hop.h, the header file for the C++ program for the Hopfield network. The include files
listed in it are the stdio.h, iostream.h, and math.h. The iostream.h file contains the declarations and details of
the C++ streams for input and output. A network class and a neuron class, are declared in Hop.h. The data
members and member functions are declared within each class, and their accessibility is specified by the
keywords protected or public.
Listing 4.1 Header file for C++ program for Hopfield network.
//Hop.h V. Rao, H. Rao

//Single layer Hopfield Network with 4 neurons
#include <stdio.h>
#include <math.h>
class neuron
{
protected:
int activation;
friend class network;
public:
int weightv[4];
neuron() {};
neuron(int *j) ;
int act(int, int*);
};
class network
{
public:
neuron nrn[4];
int output[4];
C++ Program for a Hopfield Network 56

int threshld(int) ;
void activation(int j[4]);
network(int*,int*,int*,int*);
};
Notes on the Header File Hop.h
Notice that the data item activation in the neuron class is declared as protected. In order to make the member
activation of the neuron class accessible to the network class, the network is declared a friend class in the
class neuron. Also, there are two constructors for the class neuron. One of them creates the object neuron
without initializing any data members. The other creates the object neuron and initializes the connection
weights.
Source Code for the Hopfield Network
Listing 4.2 contains the source code for the C++ program for a Hopfield network in the file Hop.cpp. The
member functions of the classes declared in Hop.h are implemented here. The function main contains the
input patterns, values to initialize the weight matrix, and calls to the constructor of network class and other
member functions of the network class.
Listing 4.2 Source code for C++ program for Hopfield network.
//Hop.cpp V. Rao, H. Rao

//Single layer Hopfield Network with 4 neurons
#include "hop.h"
neuron::neuron(int *j)
{
int i;
for(i=0;i<4;i++)
{
weightv[i]= *(j+i);
}
}
int neuron::act(int m, int *x)

{
int i;
int a=0;
for(i=0;i<m;i++)
{
a += x[i]*weightv[i];
}
return a;
}
int network::threshld(int k)
{
if(k>=0)
return (1);
else
return (0);
}
network::network(int a[4],int b[4],int c[4],int d[4])

{
Notes on the Header File Hop.h 57

nrn[0] = neuron(a) ;
nrn[1] = neuron(b) ;
nrn[2] = neuron(c) ;
nrn[3] = neuron(d) ;
}
void network::activation(int *patrn)

{
int i,j;
for(i=0;i<4;i++)
{
for(j=0;j<4;j++)
{
cout<<"\n nrn["<<i<<"].weightv["<<j<<"] is "
<<nrn[i].weightv[j];
}
nrn[i].activation = nrn[i].act(4,patrn);
cout<<"\nactivation is "<<nrn[i].activation;
output[i]=threshld(nrn[i].activation);
cout<<"\noutput value is "<<output[i]<<"\n";
}
}
void main ()
{
int patrn1[]= {1,0,1,0},i;
int wt1[]= {0,−3,3,−3};
int wt2[]= {−3,0,−3,3};
int wt3[]= {3,−3,0,−3};
int wt4[]= {−3,3,−3,0};
cout<<"\nTHIS PROGRAM IS FOR A HOPFIELD NETWORK WITH A SINGLE LAYER OF";

cout<<"\n4 FULLY INTERCONNECTED NEURONS. THE NETWORK SHOULD RECALL THE";
cout<<"\nPATTERNS 1010 AND 0101 CORRECTLY.\n";
//create the network by calling its constructor.

// the constructor calls neuron constructor as many times as the number of
// neurons in the network.
network h1(wt1,wt2,wt3,wt4);
//present a pattern to the network and get the activations of the neurons
h1.activation(patrn1);
//check if the pattern given is correctly recalled and give message

for(i=0;i<4;i++)
{
if (h1.output[i] == patrn1[i])
cout<<"\n pattern= "<<patrn1[i]<<
" output = "<<h1.output[i]<<" component matches";
else
" output = "<<h1.output[i]<<
" discrepancy occurred";
}
cout<<"\n\n";
int patrn2[]= {0,1,0,1};
h1.activation(patrn2);
for(i=0;i<4;i++)
{
if (h1.output[i] == patrn2[i])

" output = "<<h1.output[i]<<" component matches";

else
" output = "<<h1.output[i]<<
" discrepancy occurred";
}
}


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Comments on the C++ Program for Hopfield Network
Note the use of the output stream operator cout<< to output text strings or numerical output. C++ has istream
and ostream classes from which the iostream class is derived. The standard input and output streams are cin
and cout, respectively, used, correspondingly, with the operators >> and <<. Use of cout for the output stream
is much simpler than the use of the C function printf. As you can see, there is no formatting suggested for
output. However, there is a provision that allows you to format the output, while using cout.
Also note the way comments are introduced in the program. The line with comments should start with a
double slash //. Unlike C, the comment does not have to end with a double slash. Of course, if the comments
extend to subsequent lines, each such line should have a double slash at the start. You can still use the pair, /*
at the beginning with */ at the end of lines of comments, as you do in C. If the comment continues through
many lines, the C facility will be handier to delimit the comments.
The neurons in the network are members of the network class and are identified by the abbreviation nrn. The
two patterns, 1010 and 0101, are presented to the network one at a time in the program.
Output from the C++ Program for Hopfield Network
The output from this program is as follows and is self−explanatory. When you run this program, you’re likely
to see a lot of output whiz by, so in order to leisurely look at the output, use redirection. Type Hop >
filename, and your output will be stored in a file, which you can edit with any text editor or list by using the
type filename | more command.
THIS PROGRAM IS FOR A HOPFIELD NETWORK WITH A SINGLE LAYER OF 4 FULLY

INTERCONNECTED NEURONS. THE NETWORK SHOULD RECALL THE PATTERNS 1010 AND
0101 CORRECTLY.
nrn[0].weightv[0] is 0
nrn[0].weightv[1] is −3
activation is 3
output value is 1
activation is −6
output value is 0
activation is 3
Comments on the C++ Program for Hopfield Network 60

output value is 1
activation is −6
output value is 0
pattern= 1 output = 1 component matches

activation is −6
output value is 0
activation is 3
output value is 1
activation is −6
output value is 0
activation is 3
output value is 1

Further Comments on the Program and Its Output
Let us recall our previous discussion of this example in Chapter 1. What does the network give as output if we
present a pattern different from both A and B? If C = (0, 1, 0, 0) is the input pattern, the activation (dot
products) would be –3, 0, –3, 3 making the outputs (next state) of the neurons 0,1,0,1, so that B would be
recalled. This is quite interesting, because if we intended to input B, and we made a slight error and ended up
presenting C instead, the network would recall B. You can run the program by changing the pattern to 0, 1, 0,
0 and compiling again, to see that the B pattern is recalled.
Another element about the example in Chapter 1 is that the weight matrix W is not the only weight matrix that
would enable the network to recall the patterns A and B correctly. If we replace the 3 and –3 in the matrix
with 2 and –2, respectively, the resulting matrix would facilitate the same performance from the network. One
Further Comments on the Program and Its Output 61

way for you to check this is to change the wt1, wt2, wt3, wt4 given in the program accordingly, and compile
and run the program again. The reason why both of the weight matrices work is that they are closely related.
In fact, one is a scalar (constant) multiple of the other, that is, if you multiply each element in the matrix by
the same scalar, namely 2/3, you get the corresponding matrix in cases where 3 and –3 are replaced with 2 and
–2, respectively.
Further Comments on the Program and Its Output 62


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
A New Weight Matrix to Recall More Patterns

Let’s continue to discuss this example. Suppose we are interested in having the patterns E = (1, 0, 0, 1) and F
= (0, 1, 1, 0) also recalled correctly, in addition to the patterns A and B. In this case we would need to train the
network and come up with a learning algorithm, which we will discuss in more detail later in the book. We
come up with the matrix W1, which follows.
0 −5 4 4
W1 = −5 0 4 4
4 4 0 −5
4 4 −5 0
Try to use this modification of the weight matrix in the source program, and then compile and run the program
to see that the network successfully recalls all four patterns A, B, E, and F.
NOTE: The C++ implementation shown does not include the asynchronous update feature
mentioned in Chapter 1, which is not necessary for the patterns presented. The coding of this
feature is left as an exercise for the reader.
Weight Determination
You may be wondering about how these weight matrices were developed in the previous example, since so far
we’ve only discussed how the network does its job, and how to implement the model. You have learned that
the choice of weight matrix is not necessarily unique. But you want to be assured that there is some
established way besides trial and error, in which to construct a weight matrix. You can go about this in the
following way.
Binary to Bipolar Mapping

Let’s look at the previous example. You have seen that by replacing each 0 in a binary string with a –1, you
get the corresponding bipolar string. If you keep all 1’s the same and replace each 0 with a –1, you will have a
formula for the above option. You can apply the following function to each bit in the string:
f(x) = 2x – 1
NOTE: When you give the binary bit x, you get the corresponding bipolar character f(x)
For inverse mapping, which turns a bipolar string into a binary string, you use the following function:
f(x) = (x + 1) / 2
A New Weight Matrix to Recall More Patterns 63

NOTE: When you give the bipolar character x, you get the corresponding binary bit f(x)
Pattern’s Contribution to Weight
Next, we work with the bipolar versions of the input patterns. You take each pattern to be recalled, one at a
time, and determine its contribution to the weight matrix of the network. The contribution of each pattern is
itself a matrix. The size of such a matrix is the same as the weight matrix of the network. Then add these
contributions, in the way matrices are added, and you end up with the weight matrix for the network, which is
also referred to as the correlation matrix. Let us find the contribution of the pattern A = (1, 0, 1, 0):
First, we notice that the binary to bipolar mapping of A = (1, 0, 1, 0) gives the vector (1, –1, 1, –1).
Then we take the transpose, and multiply, the way matrices are multiplied, and we see the following:
1 [1 −1 1 −1] 1 −1 1 −1
1 = −1 1 −1 1
1 1 −1 1 −1
1 −1 1 −1 1
Now subtract 1 from each element in the main diagonal (that runs from top left to bottom right). This
operation gives the same result as subtracting the identity matrix from the given matrix, obtaining 0’s in the
main diagonal. The resulting matrix, which is given next, is the contribution of the pattern (1, 0, 1, 0) to the
weight matrix.
0 −1 1 −1
−1 0 −1 1
1 −1 0 −1
−1 1 −1 0
Similarly, we can calculate the contribution from the pattern B = (0, 1, 0, 1) by verifying that pattern B’s
contribution is the same matrix as pattern A’s contribution. Therefore, the matrix of weights for this exercise
is the matrix W shown here.
0 −2 2 −2
W = −2 0 −2 2
2 −2 0 −2
−2 2 −2 0
You can now optionally apply an arbitrary scalar multiplier to all the entries of the matrix if you wish. This is
how we had previously obtained the +/− 3 values instead of +/− 2 values shown above.
Pattern’s Contribution to Weight 64


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Autoassociative Network
The Hopfield network just shown has the feature that the network associates an input pattern with itself in
recall. This makes the network an autoassociative network. The patterns used for determining the proper
weight matrix are also the ones that are autoassociatively recalled. These patterns are called the exemplars. A
pattern other than an exemplar may or may not be recalled by the network. Of course, when you present the
pattern 0 0 0 0, it is stable, even though it is not an exemplar pattern.
Orthogonal Bit Patterns

You may be wondering how many patterns the network with four nodes is able to recall. Let us first consider
how many different bit patterns are orthogonal to a given bit pattern. This question really refers to bit patterns
in which at least one bit is equal to 1. A little reflection tells us that if two bit patterns are to be orthogonal,
they cannot both have 1’s in the same position, since the dot product would need to be 0. In other words, a
bitwise logical AND operation of the two bit patterns has to result in a 0. This suggests the following. If a
pattern P has k, less than 4, bit positions with 0 (and so 4−k bit positions with 1), and if pattern Q is to be
orthogonal to P, then Q can have 0 or 1 in those k positions, but it must have only 0 in the rest 4−k positions.
Since there are two choices for each of the k positions, there are 2k possible patterns orthogonal to P. This
number 2k of patterns includes the pattern with all zeroes. So there really are 2k–1 non−zero patterns
orthogonal to P. Some of these 2k–1 patterns are not orthogonal to each other. As an example, P can be the
pattern 0 1 0 0, which has k = 3 positions with 0. There are 23–1=7 nonzero patterns orthogonal to 0 1 0 0.
Among these are patterns 1 0 1 0 and 1 0 0 1, which are not orthogonal to each other, since their dot product is
1 and not 0.
Network Nodes and Input Patterns

Since our network has four neurons in it, it also has four nodes in the directed graph that represents the
network. These are laterally connected because connections are established from node to node. They are
lateral because the nodes are all in the same layer. We started with the patterns A = (1, 0, 1, 0) and B = (0, 1,
0, 1) as the exemplars. If we take any other nonzero pattern that is orthogonal to A, it will have a 1 in a
position where B also has a 1. So the new pattern will not be orthogonal to B. Therefore, the orthogonal set of
patterns that contains A and B can have only those two as its elements. If you remove B from the set, you can
get (at most) two others to join A to form an orthogonal set. They are the patterns (0, 1, 0, 0) and (0, 0, 0, 1).
If you follow the procedure described earlier to get the correlation matrix, you will get the following weight
matrix:
0 −1 3 −1
W = −1 0 −1 −1
3 −1 0 −1
−1 −1 −1 0
Autoassociative Network 65
With this matrix, pattern A is recalled, but the zero pattern (0, 0, 0, 0) is obtained for the two patterns (0, 1, 0,
0) and (0, 0, 0, 1). Once the zero pattern is obtained, its own recall will be stable.
Second Example for C++ Implementation

Recall the cash register game from the show The Price is Right, used as one of the examples in Chapter 1.
This example led to the description of the Perceptron neural network. We will now resume our discussion of
the Perceptron model and follow up with its C++ implementation. Keep the cash register game example in
mind as you read the following C++ implementation of the Perceptron model. Also note that the input signals
in this example are not necessarily binary, but they may be real numbers. It is because the prices of the items
the contestant has to choose are real numbers (dollars and cents). A Perceptron has one layer of input neurons
and one layer of output neurons. Each input layer neuron is connected to each neuron in the output layer.
Second Example for C++ Implementation 66


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
C++ Implementation of Perceptron Network
In our C++ implementation of this network, we have the following classes: we have separate classes for input
neurons and output neurons. The ineuron class is for the input neurons. This class has weight and activation
as data members. The oneuron class is similar and is for the output neuron. It is declared as a friend class in
the ineuron class. The output neuron class has also a data member called output. There is a network class,
which is a friend class in the oneuron class. An instance of the network class is created with four input
neurons. These four neurons are all connected with one output neuron.
The member functions of the ineuron class are: (1) a default constructor, (2) a second constructor that takes a
real number as an argument, and (3) a function that calculates the output of the input neuron. The constructor
taking one argument uses that argument to set the value of the weight on the connection between the input
neuron and the output neuron. The functions that determine the neuron activations and the network output are
declared public. The activations of the neurons are calculated with functions defined in the neuron classes. A
threshold value is used by a member function of the output neuron to determine if the neuron’s activation is
large enough for it to fire, giving an output of 1.
Header File
Listing 4.3 contains percept.h, the header file for the C++ program for the Perceptron network. percept.h
contains the declarations for three classes, one for input neurons, one for output neurons, and one for network.
Listing 4.3 The percept.h header file.
//percept.h V. Rao, H. Rao

// Perceptron model
#include <stdio.h>
#include <math.h>
class ineuron
{
protected:
float weight;
float activation;
friend class oneuron;
public:
ineuron() {};
ineuron(float j) ;
float act(float x);
};
class oneuron
{
protected:
int output;
C++ Implementation of Perceptron Network 67

float activation;
public:
oneuron() { };
void actvtion(float x[4], ineuron *nrn);
int outvalue(float j) ;
};
class network
{
public:
ineuron nrn[4];
oneuron onrn;
network(float,float,float,float);
};
Implementation of Functions
The network is designed to have four neurons in the input layer. Each of them is an object of class ineuron,
and these are member classes in the class network. There is one explicitly defined output neuron of the class
oneuron. The network constructor also invokes the neuron constructor for each input layer neuron in the
network by providing it with the initial weight for its connection to the neuron in the output layer. The
constructor for the output neuron is also invoked by the network constructor, at the same time initializing the
output and activation data members of the output neuron each to zero. To make sure there is access to needed
information and functions, the output neuron is declared a friend class in the class ineuron. The network is
declared as a friend class in the class oneuron.
Source Code for Perceptron Network
Listing 4.4 contains the source code in percept.cpp for the C++ implementation of the Perceptron model
previously discussed.
Listing 4.4 Source code for Perceptron model.
//percept.cpp V. Rao, H. Rao

//Perceptron model
#include "percept.h"
#include "stdio.h"
#include "stdlib.h"
ineuron::ineuron(float j)
{
weight= j;
}
float ineuron::act(float x)
{
float a;
a = x*weight;
return a;
}
void oneuron::actvtion(float *inputv, ineuron *nrn)

{
Implementation of Functions 68
int i;
activation = 0;
for(i=0;i<4;i++)
{
cout<<"\nweight for neuron "<<i+1<<" is "<<nrn[i].weight;
nrn[i].activation = nrn[i].act(inputv[i]);
cout<<" activation is "<<nrn[i].activation;
activation += nrn[i].activation;
}
cout<<"\n\nactivation is "<<activation<<"\n";
}
int oneuron::outvalue(float j)
{
if(activation>=j)
{
cout<<"\nthe output neuron activation \
exceeds the threshold value of "<<j<<"\n";
output = 1;
}
else
{
cout<<"\nthe output neuron activation \
is smaller than the threshold value of "<<j<<"\n";
output = 0;
}
cout<<" output value is "<< output;

return (output);
}
network::network(float a,float b,float c,float d)

{
nrn[0] = ineuron(a) ;
nrn[1] = ineuron(b) ;
nrn[2] = ineuron(c) ;
nrn[3] = ineuron(d) ;
onrn = oneuron();
onrn.activation = 0;
onrn.output = 0;
}
void main (int argc, char * argv[])
{
float inputv1[]= {1.95,0.27,0.69,1.25};

float wtv1[]= {2,3,3,2}, wtv2[]= {3,0,6,2};
FILE * wfile, * infile;
int num=0, vecnum=0, i;
float threshold = 7.0;
if (argc < 2)
{
cerr << "Usage: percept Weightfile Inputfile";
exit(1);
}
// open files
wfile= fopen(argv[1], "r");

infile= fopen(argv[2], "r");
if ((wfile == NULL) || (infile == NULL))
{
cout << " Can't open a file\n";
exit(1);
}
cout<<"\nTHIS PROGRAM IS FOR A PERCEPTRON NETWORK WITH AN INPUT LAYER OF";

cout<<"\n4 NEURONS, EACH CONNECTED TO THE OUTPUT NEURON.\n";
cout<<"\nTHIS EXAMPLE TAKES REAL NUMBERS AS INPUT SIGNALS\n";
//create the network by calling its constructor.

//the constructor calls neuron constructor as many times as the number of
//neurons in input layer of the network.
cout<<"please enter the number of weights/vectors \n";

cin >> vecnum;
for (i=1;i<=vecnum;i++)
{
fscanf(wfile,"%f %f %f %f\n", &wtv1[0],&wtv1[1],&wtv1[2],&wtv1[3]);
network h1(wtv1[0],wtv1[1],wtv1[2],wtv1[3]);
fscanf(infile,"%f %f %f %f \n",
&inputv1[0],&inputv1[1],&inputv1[2],&inputv1[3]);
cout<<"this is vector # " << i << "\n";
cout << "please enter a threshold value, eg 7.0\n";
cin >> threshold;
h1.onrn.actvtion(inputv1, h1.nrn);
h1.onrn.outvalue(threshold);
cout<<"\n\n";
}
fclose(wfile);
fclose(infile);
}

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Comments on Your C++ Program
Notice the use of input stream operator cin>> in the C++ program, instead of the C function scanf in several
places. The iostream class in C++ was discussed earlier in this chapter. The program works like this:
First, the network input neurons are given their connection weights, and then an input vector is presented to
the input layer. A threshold value is specified, and the output neuron does the weighted sum of its inputs,
which are the outputs of the input layer neurons. This weighted sum is the activation of the output neuron, and
it is compared with the threshold value, and the output neuron fires (output is 1) if the threshold value is not
greater than its activation. It does not fire (output is 0) if its activation is smaller than the threshold value. In
this implementation, neither supervised nor unsupervised training is incorporated.
Input/Output for percept.cpp
There are two data files used in this program. One is for setting up the weights, and the other for setting up the
input vectors. On the command line, you enter the program name followed by the weight file name and the
input file name. For this discussion (also on the accompanying disk for this book) create a file called
weight.dat, which contains the following data:
2.0 3.0 3.0 2.0

3.0 0.0 6.0 2.0
These are two weight vectors. Create also an input file called input.dat with the two data vectors below:
1.95 0.27 0.69 1.25

0.30 1.05 0.75 0.19
During the execution of the program, you are first prompted for the number of vectors that are used (in this
case, 2), then for a threshold value for the input/weight vectors (use 7.0 in both cases). You will then see the
following output. Note that the user input is in italic.
percept weight.dat input.dat
THIS PROGRAM IS FOR A PERCEPTRON NETWORK WITH AN INPUT LAYER OF 4

NEURONS, EACH CONNECTED TO THE OUTPUT NEURON.
THIS EXAMPLE TAKES REAL NUMBERS AS INPUT SIGNALS

please enter the number of weights/vectors
2
this is vector # 1
please enter a threshold value, eg 7.0
7.0
weight for neuron 1 is 2 activation is 3.9

Comments on Your C++ Program 71


activation is 9.28
the output neuron activation exceeds the threshold value of 7

output value is 1
this is vector # 2
please enter a threshold value, eg 7.0
7.0

weight for neuron 2 is 0 activation is 0
activation is 5.78
the output neuron activation is smaller than the threshold value of 7
output value is 0
Finally, try adding a data vector of (1.4, 0.6, 0.35, 0.99) to the data file. Add a weight vector of ( 2, 6, 8, 3) to
the weight file and use a threshold value of 8.25 to see the result. You can use other values to experiment also.
Network Modeling
So far, we have considered the construction of two networks, the Hopfield memory and the Perceptron. What
are other considerations (which will be discussed in more depth in the chapters to follow) that you should
keep in mind ?
Some of the considerations that go into the modeling of a neural network for an application are:
nature of inputs
fuzzy
binary
analog
crisp
binary
analog
number of inputs
nature of outputs
fuzzy
binary
analog
crisp
binary
analog
Network Modeling 72
number of outputs
nature of the application

to complete patterns (recognize corrupted patterns)
to classify patterns
to do an optimization
to do approximation
to perform data clustering
to compute functions
dynamics
adaptive
learning
training
with exemplars
without exemplars
self−organizing
nonadaptive
learning
training
with exemplars
without exemplars
self−organizing
hidden layers
number
fixed
variable
sizes
fixed
variable
processing
additive
multiplicative
hybrid
additive and multiplicative

combining other approaches
expert systems
genetic algorithms
Network Modeling 73
Hybrid models, as indicated above, could be of the variety of combining neural network approach with expert
system methods or of combining additive and multiplicative processing paradigms.
Decision support systems are amenable to approaches that combine neural networks with expert systems. An
example of a hybrid model that combines different modes of processing by neurons is the Sigma Pi neural
network, wherein one layer of neurons uses summation in aggregation and the next layer of neurons uses
multiplicative processing.
A hidden layer, if only one, in a neural network is a layer of neurons that operates in between the input layer
and the output layer of the network. Neurons in this layer receive inputs from those in the input layer and
supply their outputs as the inputs to the neurons in the output layer. When a hidden layer comes in between
other hidden layers, it receives input and supplies input to the respective hidden layers.
In modeling a network, it is often not easy to determine how many, if any, hidden layers, and of what sizes,
are needed in the model. Some approaches, like genetic algorithms—which are paradigms competing with
neural network approaches in many situations but nevertheless can be cooperative, as here—are at times used
to make a determination on the needed or optimum, as the case may be, numbers of hidden layers and/or the
neurons in those hidden layers. In what follows, we outline one such application.
Network Modeling 74

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Tic−Tac−Toe Anyone?
David Fogel describes evolutionary general problem solving and uses the familiar game of Tic−Tac−Toe as
an example. The idea is to come up with optimal strategies in playing this game. The first player’s marker is
an X, and the second player’s marker is an O. Whoever gets three of his or her markers in a row or a column
or a diagonal before the other player does, wins. Shrewd players manage a draw position, if their equally
shrewd opponent thwarts their attempts to win. A draw position is one where neither player has three of his or
her markers in a row, or a column, or a diagonal.
The board can be described by a vector of nine components, each of which is a three−valued number. Imagine
the squares of the board for the game as taken in sequence row by row from top to bottom. Allow a 1 to show
the presence of an X in that square, a 0 to indicate a blank there, and a −1 to correspond to an O. This is an
example of a coding for the status of the board. For example, (−1, 0, 1, 0, −1, 0, 1, 1, −1) is a winning position
for the second player, because it corresponds to the board looking as below.
O X
O
X X O
A neural network for this problem will have an input layer with nine neurons, as each input pattern has nine
components. There would be some hidden layers. But the example is with one hidden layer. The output layer
also contains nine neurons, so that one cycle of operation of the network shows what the best configuration of
the board is to be, given a particular input. Of course, during this cycle of operation, all that needs to be
determined is which blank space, indicated by a 0 in the input, should be changed to 1, if strategy is being
worked for player 1. None of the 1’s and −1’s is to be changed.
In this particular example, the neural network architecture itself is dynamic. The network expands or contracts
according to some rules, which are described next.
Fogel describes the network as an evolving network in the sense that the number of neurons in the hidden
layer changed with a probability of 0.5. A node was equally likely to be added or deleted. Since the number of
unmarked squares dwindles after each play, this kind of approach with varying numbers of neurons in the
network seems to be reasonable, and interesting.
The initial set of weights is random values between −0.5 and 0.5, inclusive, according to a uniform
distribution. Bias and threshold values also come from this distribution. The sigmoid function:
1/(1 + e−x)
is used also, to determine the outputs.
Weights and biases were changed during the network operation training cycles. Thus, the network had a
learning phase. (You will read more on learning in Chapter 6.) This network is adaptive, since it changes its
Tic−Tac−Toe Anyone? 75
architecture. Other forms of adaptation in neural networks are in changing parameter values for a fixed
architecture. (See Chapter 6.) The results of the experiment Fogel describes show that you need nine neurons
in the hidden layer also, for your network to be the best for this problem. They also purged any strategy that
was likely to lose.
Fogel’s emphasis is on the evolutionary aspect of an adaptive process or experiment. Our interest in this
example is primarily due to the fact that an adaptive neural network is used.
The choice of Tic−Tac−Toe, while being a simple and all too familiar game, is in the genre of much more
complicated games. These games ask a player to place a marker in some position in a given array, and as
players take turns doing so, some criterion determines if it is a draw, or who won. Unlike in Tic−Tac−Toe, the
criterion by which one wins may not be known to the players.
Stability and Plasticity

We discuss now a few other considerations in neural network modeling by introducing short−term memory
and long−term memory concepts. Neural network training is usually done in an iterative way, meaning that
the procedure is repeated a certain number of times. These iterations are referred to as cycles. After each
cycle, the input used may remain the same or change, or the weights may remain the same or change. Such
change is based on the output of a completed cycle. If the number of cycles is not preset, and the network is
allowed to go through cycles until some other criterion is met, the question of whether or not the termination
of the iterative process occurs eventually, arises naturally.
Stability and Plasticity 76


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Stability for a Neural Network
Stability refers to such convergence that facilitates an end to the iterative process. For example, if any two
consecutive cycles result in the same output for the network, then there may be no need to do more iterations.
In this case, convergence has occurred, and the network has stabilized in its operation. If weights are being
modified after each cycle, then convergence of weights would constitute stability for the network.
In some situations, it takes many more iterations than you desire, to have output in two consecutive cycles to
be the same. Then a tolerance level on the convergence criterion can be used. With a tolerance level, you
accomplish early but satisfactory termination of the operation of the network.
Plasticity for a Neural Network
Suppose a network is trained to learn some patterns, and in this process the weights are adjusted according to
an algorithm. After learning these patterns and encountering a new pattern, the network may modify the
weights in order to learn the new pattern. But what if the new weight structure is not responsive to the new
pattern? Then the network does not possess plasticity—the ability to deal satisfactorily with new short−term
memory (STM) while retaining long−term memory (LTM). Attempts to endow a network with plasticity may
have some adverse effects on the stability of your network.
Short−Term Memory and Long−Term Memory

We alluded to short−term memory (STM) and long−term memory (LTM) in the previous paragraph. STM is
basically the information that is currently and perhaps temporarily being processed. It is manifested in the
patterns that the network encounters. LTM, on the other hand, is information that is already stored and is not
being currently processed. In a neural network, STM is usually characterized by patterns and LTM is
characterized by the connections’ weights. The weights determine how an input is processed in the network to
yield output. During the cycles of operation of a network, the weights may change. After convergence, they
represent LTM, as the weight levels achieved are stable.
Summary
You saw in this chapter, the C++ implementations of a simple Hopfield network and of a simple Perceptron
network. What have not been included in them is an automatic iteration and a learning algorithm. They were
not necessary for the examples that were used in this chapter to show C++ implementation, the emphasis was
on the method of implementation. In a later chapter, you will read about the learning algorithms and examples
of how to implement some of them.
Considerations in modeling a neural network are presented in this chapter along with an outline of how
Tic−Tac−Toe is used as an example of an adaptive neural network model.
Stability for a Neural Network 77

You also were introduced to the following concepts: stability, plasticity, short−term memory, and long−term
memory (discussed further in later chapters). Much more can be said about them, in terms of the so−called
noise−saturation dilemma, or stability–plasticity dilemma and what research has developed to address them
(for further reading, see References).
Stability for a Neural Network 78


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Chapter 5
A Survey of Neural Network Models
Neural Network Models
You were introduced in the preceding pages to the Perceptron model, the Feedforward network, and the
Hopfield network. You learned that the differences between the models lie in their architecture, encoding, and
recall. We aim now to give you a comprehensive picture of these and other neural network models. We will
show details and implementations of some networks in later chapters.
The models we briefly review in this chapter are the Perceptron, Hopfield, Adaline, Feed−Forward
Backpropagation, Bidirectional Associative Memory, Brain−State−in−a−Box, Neocognitron, Fuzzy
Associative Memory, ART1, and ART2. C++ implementations of some of these and the role of fuzzy logic in
some will be treated in the subsequent chapters. For now, our discussion will be about the distinguishing
characteristics of a neural network. We will follow it with the description of some of the models.
Layers in a Neural Network

A neural network has its neurons divided into subgroups, or fields, and elements in each subgroup are placed
in a row, or a column, in the diagram depicting the network. Each subgroup is then referred to as a layer of
neurons in the network. A great many models of neural networks have two layers, quite a few have one layer,
and some have three or more layers. A number of additional, so−called hidden layers are possible in some
networks, such as the Feed−forward backpropagation network. When the network has a single layer, the input
signals are received at that layer, processing is done by its neurons, and output is generated at that layer. When
more than one layer is present, the first field is for the neurons that supply the input signals for the neurons in
the next layer.
Every network has a layer of input neurons, but in most of the networks, the sole purpose of these neurons is
to feed the input to the next layer of neurons. However, there are feedback connections, or recurrent
connections in some networks, so that the neurons in the input layer may also do some processing. In the
Hopfield network you saw earlier, the input and output layers are the same. If any layer is present between the
input and output layers, it may be referred to as a hidden layer in general, or as a layer with a special name
taken after the researcher who proposed its inclusion to achieve certain performance from the network.
Examples are the Grossberg and the Kohonen layers. The number of hidden layers is not limited except by the
scope of the problem being addressed by the neural network.
A layer is also referred to as a field. Then the different layers can be designated as field A, field B, and so on,
or shortly, FA, FB.
Chapter 5 A Survey of Neural Network Models 79

Single−Layer Network
A neural network with a single layer is also capable of processing for some important applications, such as
integrated circuit implementations or assembly line control. The most common capability of the different
models of neural networks is pattern recognition. But one network, called the Brain−State−in−a−Box, which
is a single−layer neural network, can do pattern completion. Adaline is a network with A and B fields of
neurons, but aggregation or processing of input signals is done only by the field B neurons.
The Hopfield network is a single−layer neural network. The Hopfield network makes an association between
different patterns (heteroassociation) or associates a pattern with itself (autoassociation). You may
characterize this as being able to recognize a given pattern. The idea of viewing it as a case of pattern
recognition becomes more relevant if a pattern is presented with some noise, meaning that there is some slight
deformation in the pattern, and if the network is able to relate it to the correct pattern.
The Perceptron technically has two layers, but has only one group of weights. We therefore still refer to it as a
single−layer network. The second layer consists solely of the output neuron, and the first layer consists of the
neurons that receive input(s). Also, the neurons in the same layer, the input layer in this case, are not
interconnected, that is, no connections are made between two neurons in that same layer. On the other hand, in
the Hopfield network, there is no separate output layer, and hence, it is strictly a single−layer network. In
addition, the neurons are all fully connected with one another.
Let us spend more time on the single−layer Perceptron model and discuss its limitations, and thereby motivate
the study of multilayer networks.
Single−Layer Network 80

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
XOR Function and the Perceptron

The ability of a Perceptron in evaluating functions was brought into question when Minsky and Papert proved
that a simple function like XOR (the logical function exclusive or) could not be correctly evaluated by a
Perceptron. The XOR logical function, f(A,B), is as follows:
A B
f(A,B)= XOR(A,B)
0 0
0
0 1
1
1 0
1
1 1
0
To summarize the behavior of the XOR, if both inputs are the same value, the output is 0, otherwise the output
is 1.
Minsky and Papert showed that it is impossible to come up with the proper set of weights for the neurons in
the single layer of a simple Perceptron to evaluate the XOR function. The reason for this is that such a
Perceptron, one with a single layer of neurons, requires the function to be evaluated, to be linearly separable
by means of the function values. The concept of linear separability is explained next. But let us show you first
why the simple perceptron fails to compute this function.
Since there are two arguments for the XOR function, there would be two neurons in the input layer, and since
the function’s value is one number, there would be one output neuron. Therefore, you need two weights w1
and w2 ,and a threshold value ¸. Let us now look at the conditions to be satisfied by the w’s and the ¸ so that
the outputs corresponding to given inputs would be as for the XOR function.
First the output should be 0 if inputs are 0 and 0. The activation works out as 0. To get an output of 0, you
need 0 < ¸. This is your first condition. Table 5.1 shows this and two other conditions you need, and why.
Table 5.1 Conditions on Weights

InputActivationOutputNeeded Condition
0, 0000 < ¸
1, 0w11w1 > ¸
0, 1w21w2 > ¸
1, 1w1 + w20w1 + w2< ¸
From the first three conditions, you can deduce that the sum of the two weights has to be greater than ¸, which
has to be positive itself. Line 4 is inconsistent with lines 1, 2, and 3, since line 4 requires the sum of the two
weights to be less than ¸. This affirms the contention that it is not possible to compute the XOR function with
XOR Function and the Perceptron 81

a simple perceptron.
Geometrically, the reason for this failure is that the inputs (0, 1) and (1, 0) with which you want output 1, are
situated diagonally opposite each other, when plotted as points in the plane, as shown below in a diagram of
the output (1=T, 0=F):
F T
T F
You can’t separate the T’s and the F’s with a straight line. This means that you cannot draw a line in the plane
in such a way that neither (1, 1) −>F nor (0, 0)−>F is on the same side of the line as (0, 1) −>T and (1, 0)−> T.
Linear Separability
What linearly separable means is, that a type of a linear barrier or a separator—a line in the plane, or a plane
in the three−dimensional space, or a hyperplane in higher dimensions—should exist, so that the set of inputs
that give rise to one value for the function all lie on one side of this barrier, while on the other side lie the
inputs that do not yield that value for the function. A hyperplane is a surface in a higher dimension, but with a
linear equation defining it much the same way a line in the plane and a plane in the three−dimensional space
are defined.
To make the concept a little bit clearer, consider a problem that is similar but, let us emphasize, not the same
as the XOR problem.
Imagine a cube of 1−unit length for each of its edges and lying in the positive octant in a xyz−rectangular
coordinate system with one corner at the origin. The other corners or vertices are at points with coordinates (0,
0, 1), (0, 1, 0), (0, 1, 1), (1, 0, 0), (1, 0, 1), (1, 1, 0), and (1, 1, 1). Call the origin O, and the seven points listed
as A, B, C, D, E, F, and G, respectively. Then any two faces opposite to each other are linearly separable
because you can define the separating plane as the plane halfway between these two faces and also parallel to
these two faces.
For example, consider the faces defined by the set of points O, A, B, and C and by the set of points D, E, F,
and G. They are parallel and 1 unit apart, as you can see in Figure 5.1. The separating plane for these two
faces can be seen to be one of many possible planes—any plane in between them and parallel to them. One
example, for simplicity, is the plane that passes through the points (1/2, 0, 0), (1/2, 0, 1), (1/2, 1, 0), and (1/2,
1, 1). Of course, you need only specify three of those four points because a plane is uniquely determined by
three points that are not all on the same line. So if the first set of points corresponds to a value of say, +1 for
the function, and the second set to a value of –1, then a single−layer Perceptron can determine, through some
training algorithm, the correct weights for the connections, even if you start with the weights being initially all
0.
Figure 5.1 Separating plane.
Consider the set of points O, A, F, and G. This set of points cannot be linearly separated from the other
vertices of the cube. In this case, it would be impossible for the single−layer Perceptron to determine the
Linear Separability 82
proper weights for the neurons in evaluating the type of function we have been discussing.
Linear Separability 83

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
A Second Look at the XOR Function: Multilayer Perceptron
By introducing a set of cascaded Perceptrons, you have a Perceptron network, with an input layer, middle or
hidden layer, and an output layer. You will see that the multilayer Perceptron can evaluate the XOR function
as well as other logic functions (AND, OR, MAJORITY, etc.). The absence of the separability that we talked
about earlier is overcome by having a second stage, so to speak, of connection weights.
You need two neurons in the input layer and one in the output layer. Let us put a hidden layer with two
neurons. Let w11, w12, w21, and w22, be the weights on connections from the input neurons to the hidden layer
neurons. Let v1, v2 , be the weights on the connections from the hidden layer neurons to the outout neuron.
We will select the w’s (weights) and the threshold values ¸1 , and ¸2 at the hidden layer neurons, so that the
input (0, 0) generates the output vector (0, 0), and the input vector (1, 1) generates (1, 1), while the inputs (1,
0) and (0, 1) generate (0, 1) as the hidden layer output. The inputs to the output layer neurons would be from
the set {(0, 0), (1, 1), (0, 1)}. These three vectors are separable, with (0, 0), and (1, 1) on one side of the
separating line, while (0, 1) is on the other side.
We will select the s (weights) and Ä, the threshold value at the output neuron, so as to make the inputs (0, 0)
and (1, 1) cause an output of 0 for the network, and an output of 1 is caused by the input (0, 1). The network
layout within the labels of weights and threshold values inside the nodes representing hidden layer and output
neurons is shown in Figure 5.1a. Table 5.2 gives the results of operation of this network.
Figure 5.1a Example network.
Table 5.2 Results for the Perceptron with One Hidden Layer.
InputHidden Layer ActivationsHidden Layer OutputsOutput Neuron activatonOutput of network
(0, 0)(0, 0)(0, 0)00
(1, 1)(0.3, 0.6)(1, 1)00
(0, 1)(0.15, 0.3)(0, 1)0.31
(1, 0)(0.15, 0.3)(0, 1)0.31
It is clear from Table 5.2, that the above perceptron with a hidden layer does compute the XOR function
successfully.
Note: The activation should exceed the threshold value for a neuron to fire. Where the
output of a neuron is shown to be 0, it is because the internal activation of that neuron fell
short of its threshold value.
A Second Look at the XOR Function: Multilayer Perceptron 84

Example of the Cube Revisited
Let us return to the example of the cube with vertices at the origin O, and the points labeled A, B, C, D, E, F,
and G. Suppose the set of vertices O, A, F, and G give a value of 1 for the function to be evaluated, and the
other vertices give a –1. The two sets are not linearly separable as mentioned before. A simple Perceptron
cannot evaluate this function.
Can the addition of another layer of neurons help? The answer is yes. What would be the role of this
additional layer? The answer is that it will do the final processing for the problem after the previous layer has
done some preprocessing. This can do two separations in the sense that the set of eight vertices can be
separated—or partitioned—into three separable subsets. If this partitioning can also help collect within each
subset, like vertices, meaning those that map onto the same value for the function, the network will succeed in
its task of evaluating the function when the aggregation and thresholding is done at the output neuron.
Strategy
So the strategy is first to consider the set of vertices that give a value of +1 for the function and determine the
minimum number of subsets that can be identified to be each separable from the rest of the vertices. It is
evident that since the vertices O and A lie on one edge of the cube, they can form one subset that is separable.
The other two vertices, viz., F and one for G, which correspond to the value +1 for the function, can form a
second subset that is separable, too. We need not bother with the last four vertices from the point of view of
further partitioning that subset. It is clear that one new layer of three neurons, one of which fires for the inputs
corresponding to the vertices O and A, one for F, and G, and the third for the rest, will then facilitate the
correct evaluation of the function at the output neuron.
Example of the Cube Revisited 85


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Details
Table 5.3 lists the vertices and their coordinates, together with a flag that indicates to which subset in the
partitioning the vertex belongs. Note that you can think of the action of the Multilayer Perceptron as that of
evaluating the intersection and union of linearly separable subsets.
Table 5.3 Partitioning of Vertices of a Cube

VertexCoordinatesSubset
O(0,0,0)1
A(0,0,1)1
B(0,1,0)2
C(0,1,1)2
D(1,0,0)2
E(1,0,1)2
F(1,1,0)3
G(1,1,1)3
The network, which is a two−layer Perceptron, meaning two layers of weights, has three neurons in the first
layer and one output neuron in the second layer. Remember that we are counting those layers in which the
neurons do the aggregation of the signals coming into them using the connection weights. The first layer with
the three neurons is what is generally described as the hidden layer, since the second layer is not hidden and is
at the extreme right in the layout of the neural network. Table 5.4 gives an example of the weights you can use
for the connections between the input neurons and the hidden layer neurons. There are three input neurons,
one for each coordinate of the vertex of the cube.
Table 5.4 Weights for Connections Between Input Neurons and Hidden Layer Neurons.
Input Neuron #Hidden Layer Neuron #Connection Weight
111
120.1
13−1
211
22−1
23−1
310.2
320.3
330.6
Now we give, in Table 5.5, the weights for the connections between the three hidden−layer neurons and the
output neuron.
Table 5.5 Weights for Connection Between the Hidden−Layer Neurons and the Output Neuron
Hidden Layer Neuron #Connection Weight
Details 86
10.6
30.3
30.6
It is not apparent whether or not these weights will do the job. To determine the activations of the
hidden−layer neurons, you need these weights, and you also need the threshold value at each neuron that does
processing. A hidden−layer neuron will fire, that is, will output a 1, if the weighted sum of the signals it
receives is greater than the threshold value. If the output neuron fires, the function value is taken as +1, and if
it does not fire, the function value is –1. Table 5.6 gives the threshold values. Figure 5.1b shows the neural
network with connection weights and threshold values.
Figure 5.1b Neural Network for Cube Example
Table 5.6 Threshold Values

LayerNeuronThreshold Value
hidden11.8
hidden20.05
hidden3−0.2
output10.5
Details 87

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Performance of the Perceptron
When you input the coordinates of the vertex G, which has 1 for each coordinate, the first hidden−layer
neuron aggregates these inputs and gets a value of 2.2. Since 2.2 is more than the threshold value of the first
neuron in the hidden layer, that neuron fires, and its output of 1 becomes an input to the output neuron on the
connection with weight 0.6. But you need the activations of the other hidden−layer neurons as well. Let us
describe the performance with coordinates of G as the inputs to the network. Table 5.7 describes this.
Table 5.7 Results with Coordinates of Vertex G as Input

Vertex/
CoordinatesHidden LayerWeighted SumCommentActivationContribution to OutputSum
G:1,1,112.2>1.810.6
2−0.8<0.0500
3−1.4<−0.2000.6
The weighted sum at the output neuron is 0.6, and it is greater than the threshold value 0.5. Therefore, the
output neuron fires, and at the vertex G, the function is evaluated to have a value of +1.
Table 5.8 shows the performance of the network with the rest of the vertices of the cube. You will notice that
the network computes a value of +1 at the vertices, O, A, F, and G, and a –1 at the rest.
Table 5.8 Results with Other Inputs

Hidden Layer Neuron#Weighted SumCommentActivationContribution to OutputSum
O :0, 0, 010<1.800
20<0.0500
*
30>−0.210.60.6
A :0, 0, 110.2<1.800
20.3>0.0510.3
*
30.6>−0.210.60.9
B :0, 1, 011<1.800
2−1<0.0500
3−1<−0.2000
C :0, 1, 111.2<1.800
20.2>0.0510.3
3−0.4<−0.2000.3
D :1, 0, 011<1.800
2.1>0.0510.3
3−1<−0.2000.3
E :1, 0, 111.2<1.800
20.4>0.0510.3
3−0.4<−0.2000.3
F :1, 1, 012>1.810.6
Performance of the Perceptron 88

2−0.9<0.0500
*
3−2<−0.2000.6
*
The output neuron fires, as this value is greater than 0.5 (the threshold value); the function value is +1.
Other Two−layer Networks

Many important neural network models have two layers. The Feedforward backpropagation network, in its
simplest form, is one example. Grossberg and Carpenter’s ART1 paradigm uses a two−layer network. The
Counterpropagation network has a Kohonen layer followed by a Grossberg layer. Bidirectional Associative
Memory, (BAM), Boltzman Machine, Fuzzy Associative Memory, and Temporal Associative Memory are
other two−layer networks. For autoassociation, a single−layer network could do the job, but for
heteroassociation or other such mappings, you need at least a two−layer network. We will give more details
on these models shortly.
Many Layer Networks

Kunihiko Fukushima’s Neocognitron, noted for identifying handwritten characters, is an example of a
network with several layers. Some previously mentioned networks can also be multilayered from the addition
of more hidden layers. It is also possible to combine two or more neural networks into one network by
creating appropriate connections between layers of one subnetwork to those of the others. This would
certainly create a multilayer network.
Connections Between Layers

You have already seen some difference in the way connections are made between neurons in a neural
network. In the Hopfield network, every neuron was connected to every other in the one layer that was present
in the network. In the Perceptron, neurons within the same layer were not connected with one another, but the
connections were between the neurons in one layer and those in the next layer. In the former case, the
connections are described as being lateral. In the latter case, the connections are forward and the signals are
fed forward within the network.
Two other possibilities also exist. All the neurons in any layer may have extra connections, with each neuron
connected to itself. The second possibility is that there are connections from the neurons in one layer to the
neurons in a previous layer, in which case there is both forward and backward signal feeding. This occurs, if
feedback is a feature for the network model. The type of layout for the network neurons and the type of
connections between the neurons constitute the architecture of the particular model of the neural network.
Other Two−layer Networks 89


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Instar and Outstar

Outstar and instar are terms defined by Stephen Grossberg for ways of looking at neurons in a network. A
neuron in a web of other neurons receives a large number of inputs from outside the neuron’s boundaries. This
is like an inwardly radiating star, hence, the term instar. Also, a neuron may be sending its output to many
other destinations in the network. In this way it is acting as an outstar. Every neuron is thus simultaneously
both an instar and an outstar. As an instar it receives stimuli from other parts of the network or from outside
the network. Note that the neurons in the input layer of a network primarily have connections away from them
to the neurons in the next layer, and thus behave mostly as outstars. Neurons in the output layer have many
connections coming to it and thus behave mostly as instars. A neural network performs its work through the
constant interaction of instars and outstars.
A layer of instars can constitute a competitive layer in a network. An outstar can also be described as a source
node with some associated sink nodes that the source feeds to. Grossberg identifies the source input with a
conditioned stimulus and the sink inputs with unconditioned stimuli. Robert Hecht−Nielsen’s
Counterpropagation network is a model built with instars and outstars.
Weights on Connections
Weight assignments on connections between neurons not only indicate the strength of the signal that is being
fed for aggregation but also the type of interaction between the two neurons. The type of interaction is one of
cooperation or of competition. The cooperative type is suggested by a positive weight, and the competition by
a negative weight, on the connection. The positive weight connection is meant for what is called excitation,
while the negative weight connection is termed an inhibition.
Initialization of Weights
Initializing the network weight structure is part of what is called the encoding phase of a network operation.
The encoding algorithms are several, differing by model and by application. You may have gotten the
impression that the weight matrices used in the examples discussed in detail thus far have been arbitrarily
determined; or if there is a method of setting them up, you are not told what it is.
It is possible to start with randomly chosen values for the weights and to let the weights be adjusted
appropriately as the network is run through successive iterations. This would make it easier also. For example,
under supervised training, if the error between the desired and computed output is used as a criterion in
adjusting weights, then one may as well set the initial weights to zero and let the training process take care of
the rest. The small example that follows illustrates this point.
A Small Example
Suppose you have a network with two input neurons and one output neuron, with forward connections
between the input neurons and the output neuron, as shown in Figure 5.2. The network is required to output a
Instar and Outstar 90

1 for the input patterns (1, 0) and (1, 1), and the value 0 for (0, 1) and (0, 0). There are only two connection
weights w1 and w2.
Figure 5.2 Neural network with forward connections.
Let us set initially both weights to 0, but you need a threshold function also. Let us use the following
threshold function, which is slightly different from the one used in a previous example:
1 if x > 0
f(x) =
{ 0 if x d 0
The reason for modifying this function is that if f(x) has value 1 when x = 0, then no matter what the weights
are, the output will work out to 1 with input (0, 0). This makes it impossible to get a correct computation of
any function that takes the value 0 for the arguments (0, 0).
Now we need to know by what procedure we adjust the weights. The procedure we would apply for this
example is as follows.
• If the output with input pattern (a, b) is as desired, then do not adjust the weights.
• If the output with input pattern (a, b) is smaller than what it should be, then increment each of w 1
and w2 by 1.
• If the output with input pattern (a, b) is greater than what it should be, then subtract 1 from w 1 if
the product aw1 is smaller than 1, and adjust w2 similarly.
Table 5.9 shows what takes place when we follow these procedures, and at what values the weights settle.
Table 5.9 Adjustment of Weights

stepw1w2abactivationoutputcomment
1001100desired output is 1; increment both w’s
2111121output is what it should be
4110111output is 1; it should be 0.
5subtract 1 from w2
Instar and Outstar 91


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Table 5.9 shows that the network weight vector changed from an initial vector (0, 0) to the final weight vector
(1, 0) in eight iterations. This example is not of a network for pattern matching. If you think about it, you will
realize that the network is designed to fire if the first digit in the pattern is a 1, and not otherwise. An analogy
for this kind of a problem is determining if a given image contains a specific object in a specific part of the
image, such as a dot should occur in the letter i.
If the initial weights are chosen somewhat prudently and to make some particular relevance, then the speed of
operation can be increased in the sense of convergence being achieved with fewer iterations than otherwise.
Thus, encoding algorithms are important. We now present some of the encoding algorithms.
Initializing Weights for Autoassociative Networks
Consider a network that is to associate each input pattern with itself and which gets binary patterns as inputs.
Make a bipolar mapping on the input pattern. That is, replace each 0 by –1. Call the mapped pattern the vector
x, when written as a column vector. The transpose, the same vector written as a row vector, is xT. You will get
a matrix of order the size of x when you form the product xxT. Obtain similar matrices for the other patterns
you want the network to store. Add these matrices to give you the matrix of weights to be used initially, as we
did in Chapter 4. This process can be described with the following equation:
W = ÂixixiT
Weight Initialization for Heteroassociative Networks
Consider a network that is to associate one input pattern with another pattern and that gets binary patterns as
inputs. Make a bipolar mapping on the input pattern. That is, replace each 0 by –1. Call the mapped pattern
the vector x when written as a column vector. Get a similar bipolar mapping for the corresponding associated
pattern. Call it y. You will get a matrix of size x by size y when you form the product xyT. Obtain similar
matrices for the other patterns you want the network to store. Add these matrices to give you the matrix of
weights to be used initially. The following equation restates this process:
W = ÂixiyiT
On Center, Off Surround

In one of the many interesting paradigms you encounter in neural network models and theory, is the strategy
winner takes all. Well, if there should be one winner emerging from a crowd of neurons in a particular layer,
there needs to be competition. Since everybody is for himself in such a competition, in this case every neuron
for itself, it would be necessary to have lateral connections that indicate this circumstance. The lateral
connections from any neuron to the others should have a negative weight. Or, the neuron with the highest
activation is considered the winner and only its weights are modified in the training process, leaving the
weights of others the same. Winner takes all means that only one neuron in that layer fires and the others do
not. This can happen in a hidden layer or in the output layer.
Initializing Weights for Autoassociative Networks 92

In another situation, when a particular category of input is to be identified from among several groups of
inputs, there has to be a subset of the neurons that are dedicated to seeing it happen. In this case, inhibition
increases for distant neurons, whereas excitation increases for the neighboring ones, as far as such a subset of
neurons is concerned. The phrase on center, off surround describes this phenomenon of distant inhibition and
near excitation.
Weights also are the prime components in a neural network, as they reflect on the one hand the memory stored
by the network, and on the other hand the basis for learning and training.
Inputs
You have seen that mutually orthogonal or almost orthogonal patterns are required as stable stored patterns for
the Hopfield network, which we discussed before for pattern matching. Similar restrictions are found also
with other neural networks. Sometimes it is not a restriction, but the purpose of the model makes natural a
certain type of input. Certainly, in the context of pattern classification, binary input patterns make problem
setup simpler. Binary, bipolar, and analog signals are the varieties of inputs. Networks that accept analog
signals as inputs are for continuous models, and those that require binary or bipolar inputs are for discrete
models. Binary inputs can be fed to networks for continuous models, but analog signals cannot be input to
networks for discrete models (unless they are fuzzified). With input possibilities being discrete or analog, and
the model possibilities being discrete or continuous, there are potentially four situations, but one of them
where analog inputs are considered for a discrete model is untenable.
An example of a continuous model is where a network is to adjust the angle by which the steering wheel of a
truck is to be turned to back up the truck into a parking space. If a network is supposed to recognize characters
of the alphabet, a means of discretization of a character allows the use of a discrete model.
What are the types of inputs for problems like image processing or handwriting analysis? Remembering that
artificial neurons, as processing elements, do aggregation of their inputs by using connection weights, and that
the output neuron uses a threshold function, you know that the inputs have to be numerical. A handwritten
character can be superimposed on a grid, and the input can consist of the cells in each row of the grid where a
part of the character is present. In other words, the input corresponding to one character will be a set of binary
or gray−scale sequences containing one sequence for each row of the grid. A 1 in a particular position in the
sequence for a row shows that the corresponding pixel is present(black) in that part of the grid, while 0 shows
it is not. The size of the grid has to be big enough to accommodate the largest character under study, as well as
the most complex features.
Inputs 93

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Outputs
The output from some neural networks is a spatial pattern that can include a bit pattern, in some a binary
function value, and in some others an analog signal. The type of mapping intended for the inputs determines
the type of outputs, naturally. The output could be one of classifying the input data, or finding associations
between patterns of the same dimension as the input.
The threshold functions do the final mapping of the activations of the output neurons into the network
outputs. But the outputs from a single cycle of operation of a neural network may not be the final outputs,
since you would iterate the network into further cycles of operation until you see convergence. If convergence
seems possible, but is taking an awful lot of time and effort, that is, if it is too slow to learn, you may assign a
tolerance level and settle for the network to achieve near convergence.
The Threshold Function

The output of any neuron is the result of thresholding, if any, of its internal activation, which, in turn, is the
weighted sum of the neuron’s inputs. Thresholding sometimes is done for the sake of scaling down the
activation and mapping it into a meaningful output for the problem, and sometimes for adding a bias.
Thresholding (scaling) is important for multilayer networks to preserve a meaningful range across each layer’s
operations. The most often used threshold function is the sigmoid function. A step function or a ramp
function or just a linear function can be used, as when you simply add the bias to the activation. The sigmoid
function accomplishes mapping the activation into the interval [0, 1]. The equations are given as follows for
the different threshold functions just mentioned.
The Sigmoid Function
More than one function goes by the name sigmoid function. They differ in their formulas and in their ranges.
They all have a graph similar to a stretched letter s. We give below two such functions. The first is the
hyperbolic tangent function with values in (–1, 1). The second is the logistic function and has values between
0 and 1. You therefore choose the one that fits the range you want. The graph of the sigmoid logistic function
is given in Fig. 5.3.
x
1. f(x) = tanh(x) = ( e − e−x) / (ex + e−x)
2. f(x) = 1 / (1+ e −x)
Note that the first function here, the hyperbolic tangent function, can also be written, as 1 − 2e−x / (ex + e−x )
after adding and also subtracting e−x to the numerator, and then simplifying. If now you multiply in the second
term both numerator and denominator by ex, you get 1 − 2/ (e2x + 1). As x approaches −, this function goes to
−1, and as x approaches +, it goes to +1. On the other hand, the second function here, the sigmoid logistic
function, goes to 0 as x approaches −, and to +1 as x approaches +. You can see this if you rewrite 1 / (1+
e−x) as 1 − 1 / (1+ ex), after manipulations similar to those above.
Outputs 94
You can think of equation 1 as the bipolar equivalent of binary equation 2. Both functions have the same
shape.
Figure 5.3 is the graph of the sigmoid logistic function (number 2 of the preceding list).
Figure 5.3 The sigmoid function.
The Step Function
The step function is also frequently used as a threshold function. The function is 0 to start with and remains
so to the left of some threshold value ¸. A jump to 1 occurs for the value of the function to the right of ¸, and
the function then remains at the level 1. In general, a step function can have a finite number of points at which
jumps of equal or unequal size occur. When the jumps are equal and at many points, the graph will resemble a
staircase. We are interested in a step function that goes from 0 to 1 in one step, as soon as the argument
exceeds the threshold value ¸. You could also have two values other than 0 and 1 in defining the range of
values of such a step function. A graph of the step function follows in Figure 5.4.
Figure 5.4 The step function.
Note: You can think of a sigmoid function as a fuzzy step function.
The Ramp Function
To describe the ramp function simply, first consider a step function that makes a jump from 0 to 1 at some
point. Instead of letting it take a sudden jump like that at one point, let it gradually gain in value, along a
straight line (looks like a ramp), over a finite interval reaching from an initial 0 to a final 1. Thus, you get a
ramp function. You can think of a ramp function as a piecewise linear approximation of a sigmoid. The
graph of a ramp function is illustrated in Figure 5.5.
Figure 5.5 Graph of a ramp function.
Linear Function
A linear function is a simple one given by an equation of the form:
f(x) = ±x + ²
When ± = 1, the application of this threshold function amounts to simply adding a bias equal to ² to the sum
of the inputs.
The Step Function 95

Applications
As briefly indicated before, the areas of application generally include auto− and heteroassociation, pattern
recognition, data compression, data completion, signal filtering, image processing, forecasting, handwriting
recognition, and optimization. The type of connections in the network, and the type of learning algorithm used
must be chosen appropriate to the application. For example, a network with lateral connections can do
autoassociation, while a feed−forward type can do forecasting.
Some Neural Network Models

Adaline and Madaline
Adaline is the acronym for adaptive linear element, due to Bernard Widrow and Marcian Hoff. It is similar to
a Perceptron. Inputs are real numbers in the interval [–1,+1], and learning is based on the criterion of
minimizing the average squared error. Adaline has a high capacity to store patterns. Madaline stands for many
Adalines and is a neural network that is widely used. It is composed of field A and field B neurons, and there
is one connection from each field A neuron to each field B neuron. Figure 5.6 shows a diagram of the
Madaline.
Figure 5.6 The Madaline model.
Applications 96

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Backpropagation
The Backpropagation training algorithm for training feed−forward networks was developed by Paul Werbos,
and later by Parker, and Rummelhart and McClelland. This type of network configuration is the most
common in use, due to its ease of training. It is estimated that over 80% of all neural network projects in
development use backpropagation. In backpropagation, there are two phases in its learning cycle, one to
propagate the input pattern through the network and the other to adapt the output, by changing the weights in
the network. It is the error signals that are backpropagated in the network operation to the hidden layer(s). The
portion of the error signal that a hidden−layer neuron receives in this process is an estimate of the contribution
of a particular neuron to the output error. Adjusting on this basis the weights of the connections, the squared
error, or some other metric, is reduced in each cycle and finally minimized, if possible.
Figure for Backpropagation Network
You will find in Figure 7.1 in Chapter 7, the layout of the nodes that represent the neurons in a feedforward
Backpropagation network and the connections between them. For now, you try your hand at drawing this
layout based on the following description, and compare your drawing with Figure 7.1. There are three fields
of neurons. The connections are forward and are from each neuron in a layer to every neuron in the next layer.
There are no lateral or recurrent connections. Labels on connections indicate weights. Keep in mind that the
number of neurons is not necessarily the same in different layers, and this fact should be evident in the
notation for the weights.
Bidirectional Associative Memory
Bidirectional Associative Memory, (BAM), and other models described in this section were developed by
Bart Kosko. BAM is a network with feedback connections from the output layer to the input layer. It
associates a member of the set of input patterns with a member of the set of output patterns that is the closest,
and thus it does heteroassociation. The patterns can be with binary or bipolar values. If all possible input
patterns are known, the matrix of connection weights can be determined as the sum of matrices obtained by
taking the matrix product of an input vector (as a column vector) with its transpose (written as a row vector).
The pattern obtained from the output layer in one cycle of operation is fed back at the input layer at the start of
the next cycle. The process continues until the network stabilizes on all the input patterns. The stable state so
achieved is described as resonance, a concept used in the Adaptive Resonance Theory.
You will find in Figure 8.1 in Chapter 8, the layout of the nodes that represent the neurons in a BAM network
and the connections between them. There are two fields of neurons. The network is fully connected with
feedback connections and forward connections. There are no lateral or recurrent connections.
Fuzzy Associative memories are similar to Bidirectional Associative memories, except that association is
established between fuzzy patterns. Chapter 9 deals with Fuzzy Associative memories.
Backpropagation 97
Temporal Associative Memory
Another type of associative memory is temporal associative memory. Amari, a pioneer in the field of neural
networks, constructed a Temporal Associative Memory model that has feedback connections between the
input and output layers. The forte of this model is that it can store and retrieve spatiotemporal patterns. An
example of a spatiotemporal pattern is a waveform of a speech segment.
Introduced by James Anderson and others, this network differs from the single−layer fully connected Hopfield
network in that Brain−State−in−a−Box uses what we call recurrent connections as well. Each neuron has a
connection to itself. With target patterns available, a modified Hebbian learning rule is used. The adjustment
to a connection weight is proportional to the product of the desired output and the error in the computed
output. You will see more on Hebbian learning in Chapter 6. This network is adept at noise tolerance, and it
can accomplish pattern completion. Figure 5.7 shows a Brain−State−in−a−Box network.
Figure 5.7 A Brain−State−in−a−Box, network.
What’s in a Name?
More like what’s in the box? Suppose you find the following: there is a square box and its corners are the
locations for an entity to be. The entity is not at one of the corners, but is at some point inside the box. The
next position for the entity is determined by working out the change in each coordinate of the position,
according to a weight matrix, and a squashing function. This process is repeated until the entity settles down
at some position. The choice of the weight matrix is such that when the entity reaches a corner of the square
box, its position is stable and no more movement takes place. You would perhaps guess that the entity finally
settles at the corner nearest to the initial position of it within the box. It is said that this kind of an example is
the reason for the name Brain−State−in−a−Box for the model. Its forte is that it represents linear
transformations. Some type of association of patterns can be achieved with this model. If an incomplete
pattern is associated with a completed pattern, it would be an example of autoassociation.
Counterpropagation
This is a neural network model developed by Robert Hecht−Nielsen, that has one or two additional layers
between the input and output layers. If it is one, the middle layer is a Grossberg layer with a bunch of outstars.
In the other case, a Kohonen layer, or a self−organizing layer, follows the input layer, and in turn is followed
by a Grossberg layer of outstars. The model has the distinction of considerably reducing training time. With
this model, you gain a tool that works like a look−up table.
Temporal Associative Memory 98


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Neocognitron
Compared to all other neural network models, Fukushima’s Neocognitron is more complex and ambitious. It
demonstrates the advantages of a multilayered network. The Neocognitron is one of the best models for
recognizing handwritten symbols. Many pairs of layers called the S layer, for simple layer, and C layer, for
complex layer, are used. Within each S layer are several planes containing simple cells. Similarly, there are
within each C layer, an equal number of planes containing complex cells. The input layer does not have this
arrangement and is like an input layer in any other neural network.
The number of planes of simple cells and of complex cells within a pair of S and C layers being the same,
these planes are paired, and the complex plane cells process the outputs of the simple plane cells. The simple
cells are trained so that the response of a simple cell corresponds to a specific portion of the input image. If
the same part of the image occurs with some distortion, in terms of scaling or rotation, a different set of simple
cells responds to it. The complex cells output to indicate that some simple cell they correspond to did fire.
While simple cells respond to what is in a contiguous region in the image, complex cells respond on the basis
of a larger region. As the process continues to the output layer, the C−layer component of the output layer
responds, corresponding to the entire image presented in the beginning at the input layer.
ART1 is the first model for adaptive resonance theory for neural networks developed by Gail Carpenter and
Stephen Grossberg. This theory was developed to address the stability–plasticity dilemma. The network is
supposed to be plastic enough to learn an important pattern. But at the same time it should remain stable
when, in short−term memory, it encounters some distorted versions of the same pattern.
ART1 model has A and B field neurons, a gain, and a reset as shown in Figure 5.8. There are top−down and
bottom−up connections between neurons of fields A and B. The neurons in field B have lateral connections as
well as recurrent connections. That is, every neuron in this field is connected to every other neuron in this
field, including itself, in addition to the connections to the neurons in field A. The external input (or
bottom−up signal), the top−down signal, and the gain constitute three elements of a set, of which at least two
should be a +1 for the neuron in the A field to fire. This is what is termed the two−thirds rule. Initially,
therefore, the gain would be set to +1. The idea of a single winner is also employed in the B field. The gain
would not contribute in the top−down phase; actually, it will inhibit. The two−thirds rule helps move toward
stability once resonance, or equilibrium, is obtained. A vigilance parameter Á is used to determine the
parameter reset. Vigilance parameter corresponds to what degree the resonating category can be predicted.
The part of the system that contains gain is called the attentional subsystem, whereas the rest, the part that
contains reset, is termed the orienting subsystem. The top−down activity corresponds to the orienting
subsystem, and the bottom−up activity relates to the attentional subsystem.
Neocognitron 99
Figure 5.8 The ART1 network.
In ART1, classification of an input pattern in relation to stored patterns is attempted, and if unsuccessful, a
new stored classification is generated. Training is unsupervised. There are two versions of training: slow and
fast. They differ in the extent to which the weights are given the time to reach their eventual values. Slow
training is governed by differential equations, and fast training by algebraic equations.
ART2 is the analog counterpart of ART1, which is for discrete cases. These are self−organizing neural
networks, as you can surmise from the fact that training is present but unsupervised. The ART3 model is for
recognizing a coded pattern through a parallel search, and is developed by Carpenter and Grossberg. It tries to
emulate the activities of chemical transmitters in the brain during what can be construed as a parallel search
for pattern recognition.
Summary
The basic concepts of neural network layers, connections, weights, inputs, and outputs have been discussed.
An example of how adding another layer of neurons in a network can solve a problem that could not be solved
without it is given in detail. A number of neural network models are introduced briefly. Learning and training,
which form the basis of neural network behavior has not been included here, but will be discussed in the
following chapter.
Summary 100

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Chapter 6
In the last chapter, we presented an overview of different neural network models. In this chapter, we continue
the broad discussion of neural networks with two important topics: Learning and Training. Here are key
questions that we would like to answer:
• How do neural networks learn?

• What does it mean for a network to learn ?
• What differences are there between supervised and unsupervised learning ?
• What training regimens are in common use for neural networks?
Objective of Learning
There are many varieties of neural networks. In the final analysis, as we have discussed briefly in Chapter 4
on network modeling, all neural networks do one or more of the following :
• Pattern classification
• Pattern completion
• Optimization
• Data clustering
• Approximation
• Function evaluation
A neural network, in any of the previous tasks, maps a set of inputs to a set of outputs. This nonlinear
mapping can be thought of as a multidimensional mapping surface. The objective of learning is to mold the
mapping surface according to a desired response, either with or without an explicit training process.

A network can learn when training is used, or the network can learn also in the absence of training. The
difference between supervised and unsupervised training is that, in the former case, external prototypes are
used as target outputs for specific inputs, and the network is given a learning algorithm to follow and calculate
new connection weights that bring the output closer to the target output. Unsupervised learning is the sort of
learning that takes place without a teacher. For example, when you are finding your way out of a labyrinth, no
teacher is present. You learn from the responses or events that develop as you try to feel your way through the
maze. For neural networks, in the unsupervised case, a learning algorithm may be given but target outputs are
not given. In such a case, data input to the network gets clustered together; similar input stimuli cause similar
responses.
Chapter 6 Learning and Training 101

When a neural network model is developed and an appropriate learning algorithm is proposed, it would be
based on the theory supporting the model. Since the dynamics of the operation of the neural network is under
study, the learning equations are initially formulated in terms of differential equations. After solving the
differential equations, and using any initial conditions that are available, the algorithm could be simplified to
consist of an algebraic equation for the changes in the weights. These simple forms of learning equations are
available for your neural networks.
At this point of our discussion you need to know what learning algorithms are available, and what they look
like. We will now discuss two main rules for learning—Hebbian learning, used with unsupervised learning
and the delta rule, used with supervised learning. Adaptations of these by simple modifications to suit a
particular context generate many other learning rules in use today. Following the discussion of these two
rules, we present variations for each of the two classes of learning: supervised learning and unsupervised
learning.
Hebb’s Rule
Learning algorithms are usually referred to as learning rules. The foremost such rule is due to Donald Hebb.
Hebb’s rule is a statement about how the firing of one neuron, which has a role in the determination of the
activation of another neuron, affects the first neuron’s influence on the activation of the second neuron,
especially if it is done in a repetitive manner. As a learning rule, Hebb’s observation translates into a formula
for the difference in a connection weight between two neurons from one iteration to the next, as a constant
[mu] times the product of activations of the two neurons. How a connection weight is to be modified is what
the learning rule suggests. In the case of Hebb’s rule, it is adding the quantity [mu]aiaj, where ai is the
activation of the ith neuron, and aj is the activation of the jth neuron to the connection weight between the ith
and jth neurons. The constant [mu] itself is referred to as the learning rate. The following equation using the
notation just described, states it succinctly:
[Delta]wij = [mu]aiaj
As you can see, the learning rule derived from Hebb’s rule is quite simple and is used in both simple and more
involved networks. Some modify this rule by replacing the quantity ai with its deviation from the average of
all as and, similarly, replacing aj by a corresponding quantity. Such rule variations can yield rules better suited
to different situations.
For example, the output of a neural network being the activations of its output layer neurons, the Hebbian
learning rule in the case of a perceptron takes the form of adjusting the weights by adding [mu] times the
difference between the output and the target. Sometimes a situation arises where some unlearning is required
for some neurons. In this case a reverse Hebbian rule is used in which the quantity [mu]aiaj is subtracted from
the connection weight under question, which in effect is employing a negative learning rate.
In the Hopfield network of Chapter 1, there is a single layer with all neurons fully interconnected. Suppose
each neuron’s output is either a + 1 or a – 1. If we take [mu] = 1 in the Hebbian rule, the resulting
modification of the connection weights can be described as follows: add 1 to the weight, if both neuron
outputs match, that is, both are +1 or –1. And if they do not match (meaning one of them has output +1 and
the other has –1), then subtract 1 from the weight.
Hebb’s Rule 102


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Delta Rule
The delta rule is also known as the least mean squared error rule (LMS). You first calculate the square of the
errors between the target or desired values and computed values, and then take the average to get the mean
squared error. This quantity is to be minimized. For this, realize that it is a function of the weights themselves,
since the computation of output uses them. The set of values of weights that minimizes the mean squared error
is what is needed for the next cycle of operation of the neural network. Having worked this out
mathematically, and having compared the weights thus found with the weights actually used, one determines
their difference and gives it in the delta rule, each time weights are to be updated. So the delta rule, which is
also the rule used first by Widrow and Hoff, in the context of learning in neural networks, is stated as an
equation defining the change in the weights to be affected.
Suppose you fix your attention to the weight on the connection between the ith neuron in one layer and the jth
neuron in the next layer. At time t, this weight is wij(t) . After one cycle of operation, this weight becomes
wij(t + 1). The difference between the two is wij(t + 1) − wij(t), and is denoted by [Delta]wij . The delta rule
then gives [Delta]wij as :
[Delta]wij = 2[mu]xi(desired output value – computed output value)j
Here, [mu] is the learning rate, which is positive and much smaller than 1, and xi is the ith component of the
input vector.
Supervised Learning
Supervised neural network paradigms to be discussed include :
• Perceptron
• Adaline
• Feedforward Backpropagation network
• Statistical trained networks (Boltzmann/Cauchy machines)
• Radial basis function networks
The Perceptron and the Adaline use the delta rule; the only difference is that the Perceptron has binary output,
while the Adaline has continuous valued output. The Feedforward Backpropagation network uses the
generalized delta rule, which is described next.
Generalized Delta Rule
While the delta rule uses local information on error, the generalized delta rule uses error information that is
not local. It is designed to minimize the total of the squared errors of the output neurons. In trying to achieve
this minimum, the steepest descent method, which uses the gradient of the weight surface, is used. (This is
also used in the delta rule.) For the next error calculation, the algorithm looks at the gradient of the error
Delta Rule 103

surface, which gives the direction of the largest slope on the error surface. This is used to determine the
direction to go to try to minimize the error. The algorithm chooses the negative of this gradient, which is the
direction of steepest descent. Imagine a very hilly error surface, with peaks and valleys that have a wide range
of magnitude. Imagine starting your search for minimum error at an arbitrary point. By choosing the negative
gradient on all iterations, you eventually end up at a valley. You cannot know, however, if this valley is the
global minimum or a local minimum. Getting stuck in a local minimum is one well−known potential problem
of the steepest descent method. You will see more on the generalized delta rule in the chapter on
backpropagation (Chapter 7).
Statistical Training and Simulated Annealing
The Boltzmann machine (and Cauchy machine) uses probabilities and statistical theory, along with an energy
function representing temperature. The learning is probabilistic and is called simulated annealing. At different
temperature levels, a different number of iterations in processing are used, and this constitutes an annealing
schedule. Use of probability distributions is for the goal of reaching a state of global minimum of energy.
Boltzmann distribution and Cauchy distribution are probability distributions used in this process. It is
obviously desirable to reach a global minimum, rather than settling down at a local minimum.
Figure 6.1 clarifies the distinction between a local minimum and a global minimum. In this figure you find the
graph of an energy function and points A and B. These points show that the energy levels there are smaller
than the energy levels at any point in their vicinity, so you can say they represent points of minimum energy.
The overall or global minimum, as you can see, is at point B, where the energy level is smaller than that even
at point A, so A corresponds only to a local minimum. It is desirable to get to B and not get stopped at A
itself, in the pursuit of a minimum for the energy function. If point C is reached, one would like the further
movement to be toward B and not A. Similarly, if a point near A is reached, the subsequent movement should
avoid reaching or settling at A but carry on to B. Perturbation techniques are useful for these considerations.
Figure 6.1 Local and global minima.
Clamping Probabilities
Sometimes in simulated annealing, first a subset of the neurons in the network are associated with some
inputs, and another subset of neurons are associated with some outputs, and these are clamped with
probabilities, which are not changed in the learning process. Then the rest of the network is subjected to
adjustments. Updating is not done for the clamped units in the network. This training procedure of Geoffrey
Hinton and Terrence Sejnowski provides an extension of the Boltzmann technique to more general networks.
Radial Basis−Function Networks
Although details of radial basis functions are beyond the scope of this book, it is worthwhile to contrast the
learning characteristics for this type of neural network model. Radial basis−function networks in topology
look similar to feedforward networks. Each neuron has an output to input characteristic that resembles a radial
function (for two inputs, and thus two dimensions). Specifically, the output h(x) is as follows:
h(x) = exp [ (x − u)2 / 2[sigma]2 ]
Statistical Training and Simulated Annealing 104

Here, x is the input vector, u is the mean, and [sigma] is the standard deviation of the output response curve of
the neuron. Radial basis function (RBF) networks have rapid training time (orders of magnitude faster than
backpropagation) and do not have problems with local minima as backpropagation does. RBF networks are
used with supervised training, and typically only the output layer is trained. Once training is completed, a
RBF network may be slower to use than a feedforward Backpropagation network, since more computations
are required to arrive at an output.
Statistical Training and Simulated Annealing 105


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Unsupervised Networks
Unsupervised neural network paradigms to be discussed include:
• Hopfield Memory
• Bidirectional associative memory
• Fuzzy associative memory
• Learning vector quantizer
• Kohonen self−organizing map
• ART1
Self−Organization
Unsupervised learning and self−organization are closely related. Unsupervised learning was mentioned in
Chapter 1, along with supervised learning. Training in supervised learning takes the form of external
exemplars being provided. The network has to compute the correct weights for the connections for neurons in
some layer or the other. Self−organization implies unsupervised learning. It was described as a characteristic
of a neural network model, ART1, based on adaptive resonance theory (to be covered in Chapter 10). With the
winner−take−all criterion, each neuron of field B learns a distinct classification. The winning neuron in a
layer, in this case the field B, is the one with the largest activation, and it is the only neuron in that layer that is
allowed to fire. Hence, the name winner take all.
Self−organization means self−adaptation of a neural network. Without target outputs, the closest possible
response to a given input signal is to be generated. Like inputs will cluster together. The connection weights
are modified through different iterations of network operation, and the network capable of self−organizing
creates on its own the closest possible set of outputs for the given inputs. This happens in the model in
Kohonen’s self−organizing map.
Kohonen’s Linear Vector Quantizer (LVQ) described briefly below is later extended as a self−organizing
feature map. Self−organization is also learning, but without supervision; it is a case of self−training.
Kohonen’s topology preserving maps illustrate self−organization by a neural network. In these cases, certain
subsets of output neurons respond to certain subareas of the inputs, so that the firing within one subset of
neurons indicates the presence of the corresponding subarea of the input. This is a useful paradigm in
applications such as speech recognition. The winner−take−all strategy used in ART1 also facilitates
self−organization.
Learning Vector Quantizer

Suppose the goal is the classification of input vectors. Kohonen’s Vector Quantization is a method in which
you first gather a finite number of vectors of the dimension of your input vector. Kohonen calls these
codebook vectors. You then assign groups of these codebook vectors to the different classes under the
classification you want to achieve. In other words, you make a correspondence between the codebook vectors
Unsupervised Networks 106

and classes, or, partition the set of codebook vectors by classes in your classification.
Now examine each input vector for its distance from each codebook vector, and find the nearest or closest
codebook vector to it. You identify the input vector with the class to which the codebook vector belongs.
Codebook vectors are updated during training, according to some algorithm. Such an algorithm strives to
achieve two things: (1), a codebook vector closest to the input vector is brought even closer to it; and (two), a
codebook vector indicating a different class is made more distant from the input vector.
For example, suppose (2, 6) is an input vector, and (3, 10) and (4, 9) are a pair of codebook vectors assigned
to different classes. You identify (2, 6) with the class to which (4, 9) belongs, since (4, 9) with a distance of
[radic]13 is closer to it than (3, 10) whose distance from (2, 6) is [radic]17. If you add 1 to each component of
(3, 10) and subtract 1 from each component of (4, 9), the new distances of these from (2, 6) are [radic]29 and
[radic]5, respectively. This shows that (3, 10) when changed to (4, 11) becomes more distant from your input
vector than before the change, and (4, 9) is changed to (3, 8), which is a bit closer to (2, 6) than (4, 9) is.
Training continues until all input vectors are classified. You obtain a stage where the classification for each
input vector remains the same as in the previous cycle of training. This is a process of self−organization.
The Learning Vector Quantizer (LVQ) of Kohonen is a self−organizing network. It classifies input vectors on
the basis of a set of stored or reference vectors. The B field neurons are also called grandmother cells, each of
which represents a specific class in the reference vector set. Either supervised or unsupervised learning can be
used with this network. (See Figure 6.2.)
Figure 6.2 Layout for Learning Vector Quantizer.
Unsupervised Networks 107


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Associative Memory Models and One−Shot Learning

The Hopfield memory, Bidirectional Associative memory and Fuzzy Associative memory are all
unsupervised networks that perform pattern completion, or pattern association. That is, with corrupted or
missing information, these memories are able to recall or complete an expected output. Gallant calls the
training method used in these networks as one−shot learning, since you determine the weight matrix as a
function of the completed patterns you wish to recall just once. An example of this was shown in Chapter 4
with determination of weights for the Hopfield memory.
Learning and Resonance

ART1 is the first neural network model based on adaptive resonance theory of Carpenter and Grossberg.
When you have a pair of patterns such that when one of them is input to a neural network the output turns out
to be the other pattern in the pair, and if this happens consistently in both directions, then you may describe it
as resonance. We discuss in Chapter 8 bidirectional associative memories and resonance. By the time training
is completed, and learning is through, many other pattern pairs would have been presented to the network as
well. If changes in the short−term memory do not disturb or affect the long−term memory, the network shows
adaptive resonance. The ART1 model is designed to maintain it. Note that this discussion relates largely to
stability.
Learning and Stability

Learning, convergence, and stability are matters of much interest. As learning is taking place, you want to
know if the process is going to halt at some appropriate point, which is a question of convergence. Is what is
learned stable, or will the network have to learn all over again, as each new event occurs? These questions
have their answers within a mathematical model with differential equations developed to describe a learning
algorithm. Proofs showing stability are part of the model inventor’s task. One particular tool that aids in the
process of showing convergence is the idea of state energy, or cost, to describe whether the direction the
process is taking can lead to convergence.
The Lyapunov function, discussed later in this chapter, is found to provide the right energy function, which
can be minimized during the operation of the neural network. This function has the property that the value
gets smaller with every change in the state of the system, thus assuring that a minimum will be reached
eventually. The Lyapunov function is discussed further because of its significant utility for neural network
models, but briefly because of the high level of mathematics involved. Fortunately, simple forms are derived
and put into learning algorithms for neural networks. The high−level mathematics is used in making the
proofs to show the viability of the models.
Alternatively, temperature relationships can be used, as in the case of the Boltzmann machine, or any other
well−suited cost function such as a function of distances used in the formulation of the Traveling Salesman
Problem, in which the total distance for the tour of the traveling salesman is to be minimized, can be
employed. The Traveling Salesman Problem is important and well−known. A set of cities is to be visited by
Associative Memory Models and One−Shot Learning 108

the salesman, each only once, and the aim is to devise a tour that minimizes the total distance traveled. The
search continues for an efficient algorithm for this problem. Some algorithms solve the problem in a large
number but not all of the situations. A neural network formulation can also work for the Traveling Salesman
Problem. You will see more about this in Chapter 15.
Training and Convergence

Suppose you have a criterion such as energy to be minimized or cost to be decreased, and you know the
optimum level for this criterion. If the network achieves the optimum value in a finite number of steps, then
you have convergence for the operation of the network. Or, if you are making pairwise associations of
patterns, there is the prospect of convergence if after each cycle of the network operation, the number of errors
is decreasing.
It is also possible that convergence is slow, so much so that it may seem to take forever to achieve the
convergence state. In that case, you should specify a tolerance value and require that the criterion be achieved
within that tolerance, avoiding a lot of computing time. You may also introduce a momentum parameter to
further change the weight and thereby speed up the convergence. One technique used is to add a portion of the
previous change in weight.
Instead of converging, the operation may result in oscillations. The weight structure may keep changing back
and forth; learning will never cease. Learning algorithms need to be analyzed in terms of convergence as
being an essential algorithm property.
Training and Convergence 109


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Lyapunov Function
Neural networks are dynamic systems in the learning and training phase of their operation, and convergence is
an essential feature, so it was necessary for the researchers developing the models and their learning
algorithms to find a provable criterion for convergence in a dynamic system. The Lyapunov function,
mentioned previously, turned out to be the most convenient and appropriate function. It is also referred to as
the energy function. The function decreases as the system states change. Such a function needs to be found
and watched as the network operation continues from cycle to cycle. Usually it involves a quadratic form. The
least mean squared error is an example of such a function. Lyapunov function usage assures a system stability
that cannot occur without convergence. It is convenient to have one value, that of the Lyapunov function
specifying the system behavior. For example, in the Hopfield network, the energy function is a constant times
the sum of products of outputs of different neurons and the connection weight between them. Since pairs of
neuron outputs are multiplied in each term, the entire expression is a quadratic form.
Other Training Issues

Besides the applications for which a neural network is intended, and depending on these applications, you
need to know certain aspects of the model. The length of encoding time and the length of learning time are
among the important considerations. These times could be long but should not be prohibitive. It is important
to understand how the network behaves with new inputs; some networks may need to be trained all over
again, but some tolerance for distortion in input patterns is desirable, where relevant. Restrictions on the
format of inputs should be known.
An advantage of neural networks is that they can deal with nonlinear functions better than traditional
algorithms can. The ability to store a number of patterns, or needing more and more neurons in the output
field with an increasing number of input patterns are the kind of aspects addressing the capabilities of a
network and also its limitations.
Adaptation
Sometimes neural networks are used as adaptive filters, the motivation for such an architecture being
selectivity. You want the neural network to classify each input pattern into its appropriate category. Adaptive
models involve changing of connection weights during all their operations, while nonadaptive ones do not
alter the weights after the phase of learning with exemplars. The Hopfield network is often used in modeling a
neural network for optimization problems, and the Backpropagation model is a popular choice in most other
applications. Neural network models are distinguishable sometimes by their architecture, sometimes by their
adaptive methods, and sometimes both. Methods for adaptation, where adaptation is incorporated, assume
great significance in the description and utility of a neural network model.
For adaptation, you can modify parameters in an architecture during training, such as the learning rate in the
backpropagation training method for example. A more radical approach is to modify the architecture itself
during training. New neural network paradigms change the number or layers and the number of neurons in a
Lyapunov Function 110

layer during training. These node adding or pruning algorithms are termed constructive algorithms. (See
Gallant for more details.)
Generalization Ability
The analogy for a neural network presented at the beginning of the chapter was that of a multidimensional
mapping surface that maps inputs to outputs. For each unseen input with respect to a training set, the
generalization ability of a network determines how well the mapping surface renders the new input in the
output space. A stock market forecaster must generalize well, otherwise you lose money in unseen market
conditions. The opposite of generalization is memorization. A pattern recognition system for images of
handwriting, should be able to generalize a letter A that is handwritten in several different ways by different
people. If the system memorizes, then you will not recognize the letter A in all cases, but instead will
categorize each letter A variation separately. The trick to achieve generalization is in network architecture,
design, and training methodology. You do not want to overtrain your neural network on expected outcomes,
but rather should accept a slightly worse than minimum error on your training set data. You will learn more
about generalization in Chapter 14.
Summary
Learning and training are important issues in applying neural networks. Two broad categories of network
learning are supervised and unsupervised learning. Supervised learning provides example outputs to compare
to while unsupervised learning does not. During supervised training, external prototypes are used as target
outputs and the network is given a learning algorithm to follow and calculate new connection weights that
bring the output closer to the target output. You can refer to networks using unsupervised learning as
self−organizing networks, since no external information or guidance is used in learning. Several neural
network paradigms were presented in this chapter along with their learning and training characteristics.
Generalization Ability 111


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Chapter 7
Backpropagation
Feedforward Backpropagation Network
The feedforward backpropagation network is a very popular model in neural networks. It does not have
feedback connections, but errors are backpropagated during training. Least mean squared error is used. Many
applications can be formulated for using a feedforward backpropagation network, and the methodology has
been a model for most multilayer neural networks. Errors in the output determine measures of hidden layer
output errors, which are used as a basis for adjustment of connection weights between the input and hidden
layers. Adjusting the two sets of weights between the pairs of layers and recalculating the outputs is an
iterative process that is carried on until the errors fall below a tolerance level. Learning rate parameters scale
the adjustments to weights. A momentum parameter can also be used in scaling the adjustments from a
previous iteration and adding to the adjustments in the current iteration.
Mapping
The feedforward backpropagation network maps the input vectors to output vectors. Pairs of input and output
vectors are chosen to train the network first. Once training is completed, the weights are set and the network
can be used to find outputs for new inputs. The dimension of the input vector determines the number of
neurons in the input layer, and the number of neurons in the output layer is determined by the dimension of
the outputs. If there are k neurons in the input layer and m neurons in the output layer, then this network can
make a mapping from k−dimensional space to an m−dimensional space. Of course, what that mapping is
depends on what pair of patterns or vectors are used as exemplars to train the network, which determine the
network weights. Once trained, the network gives you the image of a new input vector under this mapping.
Knowing what mapping you want the feedforward backpropagation network to be trained for implies the
dimensions of the input space and the output space, so that you can determine the numbers of neurons to have
in the input and output layers.
Layout
The architecture of a feedforward backpropagation network is shown in Figure 7.1. While there can be many
hidden layers, we will illustrate this network with only one hidden layer. Also, the number of neurons in the
input layer and that in the output layer are determined by the dimensions of the input and output patterns,
respectively. It is not easy to determine how many neurons are needed for the hidden layer. In order to avoid
cluttering the figure, we will show the layout in Figure 7.1 with five input neurons, three neurons in the
hidden layer, and four output neurons, with a few representative connections.
Figure 7.1 Layout of a feedforward backpropagation network.
Chapter 7 Backpropagation 112

The network has three fields of neurons: one for input neurons, one for hidden processing elements, and one
for the output neurons. As already stated, connections are for feed forward activity. There are connections
from every neuron in field A to every one in field B, and, in turn, from every neuron in field B to every
neuron in field C. Thus, there are two sets of weights, those figuring in the activations of hidden layer
neurons, and those that help determine the output neuron activations. In training, all of these weights are
adjusted by considering what can be called a cost function in terms of the error in the computed output pattern
and the desired output pattern.
Training
The feedforward backpropagation network undergoes supervised training, with a finite number of pattern
pairs consisting of an input pattern and a desired or target output pattern. An input pattern is presented at the
input layer. The neurons here pass the pattern activations to the next layer neurons, which are in a hidden
layer. The outputs of the hidden layer neurons are obtained by using perhaps a bias, and also a threshold
function with the activations determined by the weights and the inputs. These hidden layer outputs become
inputs to the output neurons, which process the inputs using an optional bias and a threshold function. The
final output of the network is determined by the activations from the output layer.
The computed pattern and the input pattern are compared, a function of this error for each component of the
pattern is determined, and adjustment to weights of connections between the hidden layer and the output layer
is computed. A similar computation, still based on the error in the output, is made for the connection weights
between the input and hidden layers. The procedure is repeated with each pattern pair assigned for training the
network. Each pass through all the training patterns is called a cycle or an epoch. The process is then repeated
as many cycles as needed until the error is within a prescribed tolerance.
There can be more than one learning rate parameter used in training in a feedforward
backpropagation network. You can use one with each set of weights between consecutive
layers.
Training 113

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Illustration: Adjustment of Weights of Connections from a Neuron in the

Hidden Layer
We will be as specific as is needed to make the computations clear. First recall that the activation of a neuron
in a layer other than the input layer is the sum of products of its inputs and the weights corresponding to the
connections that bring in those inputs. Let us discuss the jth neuron in the hidden layer. Let us be specific and
say j = 2. Suppose that the input pattern is (1.1, 2.4, 3.2, 5.1, 3.9) and the target output pattern is (0.52, 0.25,
0.75, 0.97). Let the weights be given for the second hidden layer neuron by the vector (–0.33, 0.07, –0.45,
0.13, 0.37). The activation will be the quantity:
(−0.33 * 1.1) + (0.07 * 2.4) + (−0.45 * 3.2) + (0.13 * 5.1)

+ (0.37 * 3.9) = 0.471
Now add to this an optional bias of, say, 0.679, to give 1.15. If we use the sigmoid function given by:
1 / ( 1+ exp(−x) ),
with x = 1.15, we get the output of this hidden layer neuron as 0.7595.
We are taking values to a few decimal places only for illustration, unlike the precision that
can be obtained on a computer.
We need the computed output pattern also. Let us say it turns out to be actual =(0.61, 0.41, 0.57, 0.53), while
the desired pattern is desired =(0.52, 0.25, 0.75, 0.97). Obviously, there is a discrepancy between what is
desired and what is computed. The component−wise differences are given in the vector, desired − actual =
(−0.09, −0.16, 0.18, 0.44). We use these to form another vector where each component is a product of the
error component, corresponding computed pattern component, and the complement of the latter with respect
to 1. For example, for the first component, error is –0.09, computed pattern component is 0.61, and its
complement is 0.39. Multiplying these together (0.61*0.39*−0.09), we get −0.02. Calculating the other
components similarly, we get the vector (–0.02, –0.04, 0.04, 0.11). The desired–actual vector, which is the
error vector multiplied by the actual output vector, gives you a value of error reflected back at the output of
the hidden layer. This is scaled by a value of (1−output vector), which is the first derivative of the output
activation function for numerical stability). You will see the formulas for this process later in this chapter.
The backpropagation of errors needs to be carried further. We need now the weights on the connections
between the second neuron in the hidden layer that we are concentrating on, and the different output neurons.
Let us say these weights are given by the vector (0.85, 0.62, –0.10, 0.21). The error of the second neuron in
the hidden layer is now calculated as below, using its output.
error = 0.7595 * (1 − 0.7595) * ( (0.85 * −0.02) + (0.62 * −0.04)

+ ( −0.10 * 0.04) + (0.21 * 0.11)) = −0.0041.
Illustration: Adjustment of Weights of Connections from a Neuron in the Hidden Layer 114
Again, here we multiply the error (e.g., −0.02) from the output of the current layer, by the output value
(0.7595) and the value (1−0.7595). We use the weights on the connections between neurons to work
backwards through the network.
Next, we need the learning rate parameter for this layer; let us set it as 0.2. We multiply this by the output of
the second neuron in the hidden layer, to get 0.1519. Each of the components of the vector (–0.02, –0.04, 0.04,
0.11) is multiplied now by 0.1519, which our latest computation gave. The result is a vector that gives the
adjustments to the weights on the connections that go from the second neuron in the hidden layer to the output
neurons. These values are given in the vector (–0.003, –0.006, 0.006, 0.017). After these adjustments are
added, the weights to be used in the next cycle on the connections between the second neuron in the hidden
layer and the output neurons become those in the vector (0.847, 0.614, –0.094, 0.227).
Illustration: Adjustment of Weights of Connections from a Neuron in the

Input Layer
Let us look at how adjustments are calculated for the weights on connections going from the ith neuron in the
input layer to neurons in the hidden layer. Let us take specifically i = 3, for illustration.
Much of the information we need is already obtained in the previous discussion for the second hidden layer
neuron. We have the errors in the computed output as the vector (–0.09, –0.16, 0.18, 0.44), and we obtained
the error for the second neuron in the hidden layer as –0.0041, which was not used above. Just as the error in
the output is propagated back to assign errors for the neurons in the hidden layer, those errors can be
propagated to the input layer neurons.
To determine the adjustments for the weights on connections between the input and hidden layers, we need
the errors determined for the outputs of hidden layer neurons, a learning rate parameter, and the activations of
the input neurons, which are just the input values for the input layer. Let us take the learning rate parameter to
be 0.15. Then the weight adjustments for the connections from the third input neuron to the hidden layer
neurons are obtained by multiplying the particular hidden layer neuron’s output error by the learning rate
parameter and by the input component from the input neuron. The adjustment for the weight on the
connection from the third input neuron to the second hidden layer neuron is 0.15 * 3.2 * –0.0041, which
works out to –0.002.
If the weight on this connection is, say, –0.45, then adding the adjustment of −0.002, we get the modified
weight of –0.452, to be used in the next iteration of the network operation. Similar calculations are made to
modify all other weights as well.
Illustration: Adjustment of Weights of Connections from a Neuron in the Input Layer 115

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Adjustments to Threshold Values or Biases

The bias or the threshold value we added to the activation, before applying the threshold function to get the
output of a neuron, will also be adjusted based on the error being propagated back. The needed values for this
are in the previous discussion.
The adjustment for the threshold value of a neuron in the output layer is obtained by multiplying the
calculated error (not just the difference) in the output at the output neuron and the learning rate parameter used
in the adjustment calculation for weights at this layer. In our previous example, we have the learning rate
parameter as 0.2, and the error vector as (–0.02, –0.04, 0.04, 0.11), so the adjustments to the threshold values
of the four output neurons are given by the vector (–0.004, –0.008, 0.008, 0.022). These adjustments are
added to the current levels of threshold values at the output neurons.
The adjustment to the threshold value of a neuron in the hidden layer is obtained similarly by multiplying the
learning rate with the computed error in the output of the hidden layer neuron. Therefore, for the second
neuron in the hidden layer, the adjustment to its threshold value is calculated as 0.15 * –0.0041, which is
–0.0006. Add this to the current threshold value of 0.679 to get 0.6784, which is to be used for this neuron in
the next training pattern for the neural network.
Another Example of Backpropagation Calculations

You have seen, in the preceding sections, the details of calculations for one particular neuron in the hidden
layer in a feedforward backpropagation network with five input neurons and four neurons in the output layer,
and two neurons in the hidden layer.
You are going to see all the calculations in the C++ implementation later in this chapter. Right now, though,
we present another example and give the complete picture of the calculations done in one completed iteration
or cycle of backpropagation.
Consider a feedforward backpropagation network with three input neurons, two neurons in the hidden layer,
and three output neurons. The weights on connections from the input neurons to the neurons in the hidden
layer are given in Matrix M−1, and those from the neurons in the hidden layer to output neurons are given in
Matrix M−2.
We calculate the output of each neuron in the hidden and output layers as follows. We add a bias or threshold
value to the activation of a neuron (call this result x) and use the sigmoid function below to get the output.
f(x) = 1/ (1 + e−x )
Learning parameters used are 0.2 for the connections between the hidden layer neurons and output neurons
and 0.15 for the connections between the input neurons and the neurons in the hidden layer. These values as
you recall are the same as in the previous illustration, to make it easy for you to follow the calculations by
Adjustments to Threshold Values or Biases 116

comparing them with similar calculations in the preceding sections.
The input pattern is ( 0.52, 0.75, 0.97 ), and the desired output pattern is ( 0.24, 0.17, 0.65). The initial weight
matrices are as follows:
M−1 Matrix of weights from input layer to hidden layer
0.6 − 0.4
0.2 0.8
− 0.5 0.3
M−2 Matrix of weights from hidden layer to output layer
−0.90 0.43 0.25

0.11 − 0.67 − 0.75
The threshold values (or bias) for neurons in the hidden layer are 0.2 and 0.3, while those for the output
neurons are 0.15, 0.25, and 0.05, respectively.
Table 7.1 presents all the results of calculations done in the first iteration. You will see modified or new
weight matrices and threshold values. You will use these and the original input vector and the desired output
vector to carry out the next iteration.
Table 7.1 Backpropagation Calculations

ItemI−1I−2I−3H−1H−2O−1O−2O−3
Input0.520.750.97
Desired Output0.240.170.65
M−1 Row 10.6− 0.4
M−1 Row 20.20.8
M−1 Row 3− 0.50.3
M−2 Row 1− 0.900.430.25
M−2 Row 20.11− 0.67− 0.75
Threshold0.20.30.150.250.05
Activation − H− 0.0230.683
Activation + Threshold −H0.1770.983
Output −H0.5440.728
Complement0.4560.272
Activation −O− 0.410− 0.254− 0.410
Activation + Threshold −O− 0.260− 0.004− 0.360
Output −O0.4350.4990.411
Complement0.5650.5010.589
Diff. from Target− 0.195− 0.3290.239
Computed Error −O− 0.048− 0.0820.058
Computed Error −H0.00560.0012
Adjustment to Threshold0.00080.0002−0.0096−0.01640.0116
Adjustment to M−2 Column 1−0.0005−0.0070
Adjustment to M−2 Column 20.00070.0008
Adjustment to M−2 Column 30.00080.0011
New Matrix M−2 Row 1− 0.910.4120.262
New Matrix M−2 Row 20.096− 0.694− 0.734
New Threshold Values −O0.14040.23360.0616
Adjustment to M−1 Row 10.0004−0.0001

Adjustment to M−1 Row 20.00060.0001

Adjustment to M−1 Row 30.00080.0002
New Matrix M−1 Row 10.6004− 0.4
New Matrix M−1 Row 20.20060.8001
New Matrix M−1 Row 3−0.49920.3002
New Threshold Values −H0.20080.3002


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
The top row in the table gives headings for the columns. They are, Item, I−1, I−2, I−3 (I−k being for input
layer neuron k); H−1, H−2 (for hidden layer neurons); and O−1, O−2, O−3 (for output layer neurons).
In the first column of the table, M−1 and M−2 refer to weight matrices as above. Where an entry is appended
with −H, like in Output −H, the information refers to the hidden layer. Similarly, −O refers to the output layer,
as in Activation + threshold −O.
The next iteration uses the following information from the previous iteration, which you can identify from
Table 7.1. The input pattern is ( 0.52, 0.75, 0.97 ), and the desired output pattern is ( 0.24, 0.17, 0.65). The
current weight matrices are as follows:
M−1 Matrix of weights from input layer to hidden layer:
0.6004 − 0.4
0.2006 0.8001
− 0.4992 0.3002
M−2 Matrix of weights from hidden layer to output layer:
−0.910 0.412 0.262

0.096 −0.694 −0.734
The threshold values (or bias) for neurons in the hidden layer are 0.2008 and 0.3002, while those for the
output neurons are 0.1404, 0.2336, and 0.0616, respectively.
You can keep the learning parameters as 0.15 for connections between input and hidden layer neurons, and
0.2 for connections between the hidden layer neurons and output neurons, or you can slightly modify them.
Whether or not to change these two parameters is a decision that can be made perhaps at a later iteration,
having obtained a sense of how the process is converging.
If you are satisfied with the rate at which the computed output pattern is getting close to the target output
pattern, you would not change these learning rates. If you feel the convergence is much slower than you
would like, then the learning rate parameters can be adjusted slightly upwards. It is a subjective decision both
in terms of when (if at all) and to what new levels these parameters need to be revised.
Notation and Equations

You have just seen an example of the process of training in the feedforward backpropagation network,
described in relation to one hidden layer neuron and one input neuron. There were a few vectors that were
shown and used, but perhaps not made easily identifiable. We therefore introduce some notation and describe
the equations that were implicitly used in the example.
Notation and Equations 119

Notation
Let us talk about two matrices whose elements are the weights on connections. One matrix refers to the
interface between the input and hidden layers, and the second refers to that between the hidden layer and the
output layer. Since connections exist from each neuron in one layer to every neuron in the next layer, there is
a vector of weights on the connections going out from any one neuron. Putting this vector into a row of the
matrix, we get as many rows as there are neurons from which connections are established.
Let M1 and M2 be these matrices of weights. Then what does M1[i][j] represent? It is the weight on the
connection from the ith input neuron to the jth neuron in the hidden layer. Similarly, M2[i][j] denotes the
weight on the connection from the ith neuron in the hidden layer and the jth output neuron.
Next, we will use x, y, z for the outputs of neurons in the input layer, hidden layer, and output layer,
respectively, with a subscript attached to denote which neuron in a given layer we are referring to. Let P
denote the desired output pattern, with pi as the components. Let m be the number of input neurons, so that
according to our notation, (x1, x2, …, xm) will denote the input pattern. If P has, say, r components, the
output layer needs r neurons. Let the number of hidden layer neurons be n. Let ²h be the learning rate
parameter for the hidden layer, and ²o2, that for the output layer. Let ¸ with the appropriate subscript represent
the threshold value or bias for a hidden layer neuron, and Ä with an appropriate subscript refer to the
threshold value of an output neuron.
Let the errors in output at the output layer be denoted by ejs and those at the hidden layer by ti’s. If we use a ”
prefix of any parameter, then we are looking at the change in or adjustment to that parameter. Also, the
thresholding function we would use is the sigmoid function, f(x) = 1 / (1 + exp(–x)).
Notation 120

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Equations
Output of jth hidden layer neuron:
yj = f( (£i xiM1[ i ][ j ] ) + ¸j )
(7.1)
Output of jth output layer neuron:
zj = f( (£i yiM2[ i ][ j ] ) + Äj )
(7.2)
Ith component of vector of output differences:
desired value − computed value = Pi – zi
Ith component of output error at the output layer:
ei = ( Pi − zi)
(7.3)
Ith component of output error at the hidden layer:
ti = yi (1 − yi ) (£j M2[ i ][ j ] ej)
(7.4)
Adjustment for weight between ith neuron in hidden layer and jth output neuron:
”M2[ i ][ j ] = ²o yiej
(7.5)
Adjustment for weight between ith input neuron and jth neuron in hidden layer:
M1[ i ][ j ] = ²hxitj
(7.6)
Equations 121
Adjustment to the threshold value or bias for the jth output neuron:
” ¸j = ²o ej
Adjustment to the threshold value or bias for the jth hidden layer neuron:
´¸j = ²h ej
For use of momentum parameter ± (more on this parameter in Chapter 13), instead of equations 7.5 and 7.6,
use:
”M2[ i ][ j ] ( t ) = ²o yiej + ±”M2[ i ][ j ] ( t − 1 )
(7.7)
and
”M1[ i ][ j ] ( t ) = ²h xitj + ±”M1[ i ][ j ] (t − 1)
(7.8)
C++ Implementation of a Backpropagation Simulator

The backpropagation simulator of this chapter has the following design objectives:
1. Allow the user to specify the number and size of all layers.
2. Allow the use of one or more hidden layers.
3. Be able to save and restore the state of the network.
4. Run from an arbitrarily large training data set or test data set.
5. Query the user for key network and simulation parameters.
6. Display key information at the end of the simulation.
7. Demonstrate the use of some C++ features.
A Brief Tour of How to Use the Simulator
In order to understand the C++ code, let us have an overview of the functioning of the program.
There are two modes of operation in the simulator. The user is queried first for which mode of operation is
desired. The modes are Training mode and Nontraining mode (Test mode).
Training Mode
Here, the user provides a training file in the current directory called training.dat. This file contains exemplar
pairs, or patterns. Each pattern has a set of inputs followed by a set of outputs. Each value is separated by one
or more spaces. As a convention, you can use a few extra spaces to separate the inputs from the outputs. Here
is an example of a training.dat file that contains two patterns:
0.4 0.5 0.89 −0.4 −0.8

0.23 0.8 −0.3 0.6 0.34
C++ Implementation of a Backpropagation Simulator 122

In this example, the first pattern has inputs 0.4, 0.5, and 0.89, with an expected output of –0.4 and –0.8. The
second pattern has inputs of 0.23, 0.8, and –0.3 and outputs of 0.6 and 0.34. Since there are three inputs and
two outputs, the input layer size for the network must be three neurons and the output layer size must be two
neurons. Another file that is used in training is the weights file. Once the simulator reaches the error tolerance
that was specified by the user, or the maximum number of iterations, the simulator saves the state of the
network, by saving all of its weights in a file called weights.dat. This file can then be used subsequently in
another run of the simulator in Nontraining mode. To provide some idea of how the network has done,
information about the total and average error is presented at the end of the simulation. In addition, the output
generated by the network for the last pattern vector is provided in an output file called output.dat.
Nontraining Mode (Test Mode)

In this mode, the user provides test data to the simulator in a file called test.dat. This file contains only input
patterns. When this file is applied to an already trained network, an output.dat file is generated, which
contains the outputs from the network for all of the input patterns. The network goes through one cycle of
operation in this mode, covering all the patterns in the test data file. To start up the network, the weights file,
weights.dat is read to initialize the state of the network. The user must provide the same network size
parameters used to train the network.


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Operation
The first thing to do with your simulator is to train a network with an architecture you choose. You can select
the number of layers and the number of hidden layers for your network. Keep in mind that the input and
output layer sizes are dictated by the input patterns you are presenting to the network and the outputs you seek
from the network. Once you decide on an architecture, perhaps a simple three−layer network with one hidden
layer, you prepare training data for it and save the data in the training.dat file. After this you are ready to train.
You provide the simulator with the following information:
• The mode (select 1 for training)

• The values for the error tolerance and the learning rate parameter, lambda or beta
• The maximum number of cycles, or passes through the training data you’d like to try
• The number of layers (between three and five, three implies one hidden layer, while five implies
three hidden layers)
• The size for each layer, from the input to the output
The simulator then begins training and reports the current cycle number and the average error for each cycle.
You should watch the error to see that it is on the whole decreasing with time. If it is not, you should restart
the simulation, because this will start with a brand new set of random weights and give you another, possibly
better, solution. Note that there will be legitimate periods where the error may increase for some time. Once
the simulation is done you will see information about the number of cycles and patterns used, and the total
and average error that resulted. The weights are saved in the weights.dat file. You can rename this file to use
this particular state of the network later. You can infer the size and number of layers from the information in
this file, as will be shown in the next section for the weights.dat file format. You can have a peek at the
output.dat file to see the kind of training result you have achieved. To get a full−blown accounting of each
pattern and the match to that pattern, copy the training file to the test file and delete the output information
from it. You can then run Test mode to get a full list of all the input stimuli and responses in the output.dat
file.
Summary of Files Used in the Backpropagation Simulator

Here is a list of the files for your reference, as well as what they are used for.
• weights.dat You can look at this file to see the weights for the network. It shows the layer number
followed by the weights that feed into the layer. The first layer, or input layer, layer zero, does not
have any weights associated with it. An example of the weights.dat file is shown as follows for a
network with three layers of sizes 3, 5, and 2. Note that the row width for layer n matches the column
length for layer n + 1:
1 −0.199660 −0.859660 −0.339660 −0.25966 0.520340
1 0.292860 −0.487140 0.212860 −0.967140 −0.427140
1 0.542106 −0.177894 0.322106 −0.977894 0.562106
2 −0.175350 −0.835350
2 −0.330167 −0.250167

2 0.503317 0.283317
2 −0.477158 0.222842
2 −0.928322 −0.388322
In this weights file the row width for layer 1 is 5, corresponding to the output of that (middle) layer.
The input for the layer is the column length, which is 3, just as specified. For layer 2, the output size
is the row width, which is 2, and the input size is the column length, 5, which is the same as the output
for the middle layer. You can read the weights file to find out how things look.
• training.dat This file contains the input patterns for training. You can have as large a file as
you’d like without degrading the performance of the simulator. The simulator caches data in memory
for processing. This is to improve the speed of the simulation since disk accesses are expensive in
time. A data buffer, which has a maximum size specified in a #define statement in the program, is
filled with data from the training.dat file whenever data is needed. The format for the training.dat file
has been shown in the Training mode section.
• test.dat The test.dat file is just like the training.dat file but without expected outputs. You use this
file with a trained neural network in Test mode to see what responses you get for untrained data.
• output.dat The output.dat file contains the results of the simulation. In Test mode, the input and
output vectors are shown for all pattern vectors. In the Simulator mode, the expected output is also
shown, but only the last vector in the training set is presented, since the training set is usually quite
large.
Shown here is an example of an output file in Training mode:
for input vector:
0.400000 −0.400000
output vector is:
0.880095
expected output vector is:
0.900000


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
C++ Classes and Class Hierarchy
So far, you have learned how we address most of the objectives outlined for this program. The only objective
left involves the demonstration of some C++ features. In this program we use a class hierarchy with the
inheritance feature. Also, we use polymorphism with dynamic binding and function overloading with static
binding.
First let us look at the class hierarchy used for this program (see Figure 7.2). An abstract class is a class that is
never meant to be instantiated as an object, but serves as a base class from which others can inherit
functionality and interface definitions. The layer class is such a class. You will see shortly that one of its
functions is set = zero, which indicates that this class is an abstract base class. From the layer class are two
branches. One is the input_layer class, and the other is the output_layer class. The middle layer class is very
much like the output layer in function and so inherits from the output_layer class.
Figure 7.2 Class hierarchy used in the backpropagation simulator.
Function overloading can be seen in the definition of the calc_error() function. It is used in the input_layer
with no parameters, while it is used in the output_layer (which the input_layer inherits from) with one
parameter. Using the same function name is not a problem, and this is referred to as overloading. Besides
function overloading, you may also have operator overloading, which is using an operator that performs some
familiar function like + for addition, for another function, say, vector addition.
When you have overloading with the same parameters and the keyword virtual, then you have the potential
for dynamic binding, which means that you determine which overloaded function to execute at run time and
not at compile time. Compile time binding is referred to as static binding. If you put a bunch of C++ objects in
an array of pointers to the base class, and then go through a loop that indexes each pointer and executes an
overloaded virtual function that pointer is pointing to, then you will be using dynamic binding. This is exactly
the case in the function calc_out(), which is declared with the virtual keyword in the layer base class. Each
descendant of layer can provide a version of calc_out(), which differs in functionality from the base class, and
the correct function will be selected at run time based on the object’s identity. In this case calc_out(), which is
a function to calculate the outputs for each layer, is different for the input layer than for the other two types of
layers.
Let’s look at some details in the header file in Listing 7.1:
Listing 7.1 Header file for the backpropagation simulator
// layer.h V.Rao, H. Rao

// header file for the layer class hierarchy and
// the network class
C++ Classes and Class Hierarchy 126

#define MAX_LAYERS 5
#define MAX_VECTORS 100
class network;
class layer
{
protected:
int num_inputs;
int num_outputs;
float *outputs;// pointer to array of outputs
float *inputs; // pointer to array of inputs, which
// are outputs of some other layer
friend network;
public:
virtual void calc_out()=0;
};
class input_layer: public layer

{
private:
public:
input_layer(int, int);
~input_layer();
virtual void calc_out();
};
class middle_layer;
class output_layer: public layer

{
protected:
float * weights;
float * output_errors; // array of errors at output
float * back_errors; // array of errors back−propagated
float * expected_values; // to inputs
friend network;
public:
output_layer(int, int);
~output_layer();
void calc_error(float &);
void randomize_weights();
void update_weights(const float);
void list_weights();
void write_weights(int, FILE *);
void read_weights(int, FILE *);
void list_errors();
void list_outputs();

};
class middle_layer: public output_layer

{
private:
public:
middle_layer(int, int);
~middle_layer();
void calc_error();
};
class network
private:
layer *layer_ptr[MAX_LAYERS];
int number_of_layers;
int layer_size[MAX_LAYERS];
float *buffer;
fpos_t position;
unsigned training;
public:
network();
~network();
void set_training(const unsigned &);
unsigned get_training_value();
void get_layer_info();
void set_up_network();
void write_weights(FILE *);
void read_weights(FILE *);
void write_outputs(FILE *);
void list_errors();
void forward_prop();
void backward_prop(float &);
int fill_IObuffer(FILE *);
void set_up_pattern(int);
};


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Details of the Backpropagation Header File

At the top of the file, there are two #define statements, which are used to set the maximum number of layers
that can be used, currently five, and the maximum number of training or test vectors that can be read into an
I/O buffer. This is currently 100. You can increase the size of the buffer for better speed at the cost of
increased memory usage.
The following are definitions in the layer base class. Note that the number of inputs and outputs are protected
data members, which means that they can be accessed freely by descendants of the class.
int num_inputs;
int num_outputs;
float *outputs; // pointer to array of outputs
friend network;
There are also two pointers to arrays of floats in this class. They are the pointers to the outputs in a given layer
and the inputs to a given layer. To get a better idea of what a layer encompasses, Figure 7.3 shows you a small
feedforward backpropagation network, with a dotted line that shows you the three layers for that network. A
layer contains neurons and weights. The layer is responsible for calculating its output (calc_out()), stored in
the float * outputs array, and errors (calc_error()) for each of its respective neurons. The errors are stored in
another array called float * output_errors defined in the output class. Note that the input class does not
have any weights associated with it and therefore is a special case. It does not need to provide any data
members or function members related to errors or backpropagation. The only purpose of the input layer is to
store data to be forward propagated to the next layer.
Figure 7.3 Organization of layers for backpropagation program.
With the output layer, there are a few more arrays present. First, for storing backpropagated errors, there is an
array called float * back_errors. There is a weights array called float * weights, and finally, for storing the
expected values that initiate the error calculation process, there is an array called float * expected_values.
Note that the middle layer needs almost all of these arrays and inherits them by being a derived class of the
output_layer class.
There is one other class besides the layer class and its descendants defined in this header file, and that is the
network class, which is used to set up communication channels between layers and to feed and remove data
from the network. The network class performs the interconnection of layers by setting the pointer of an input
array of a given layer to the output array of a previous layer.

This is a fairly extensible scheme that can be used to create variations on the feedforward
backpropagation network with feedback connections, for instance.
Another connection that the network class is responsible for is setting the pointer of an output_error array to
the back_error array of the next layer (remember, errors flow in reverse, and the back_error array is the
output error of the layer reflected at its inputs).
The network class stores an array of pointers to layers and an array of layer sizes for all the layers defined.
These layer objects and arrays are dynamically allocated on the heap with the New and Delete functions in
C++. There is some minimal error checking for file I/O and memory allocation, which can be enhanced, if
desired.
As you can see, the feedforward backpropagation network can quickly become a memory and CPU hog, with
large networks and large training sets. The size and topology of the network, or architecture, will largely
dictate both these characteristics.
Details of the Backpropagation Implementation File

The implementation of the classes and methods is the next topic. Let’s look at the layer.cpp file in Listing 7.2.
Listing 7.2 layer.cpp implementation file for the backpropagation simulator
// layer.cpp V.Rao, H.Rao

// compile for floating point hardware if available
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <time.h>
#include "layer.h"
inline float squash(float input)

// squashing function
// use sigmoid −− can customize to something
// else if desired; can add a bias term too
//
{
if (input < −50)
return 0.0;
else if (input > 50)
return 1.0;
else return (float)(1/(1+exp(−(double)input)));
}
inline float randomweight(unsigned init)

{
int num;
// will return a floating point
// value between −1 and 1
if (init==1) // seed the generator

num=rand() % 100;

return 2*(float(num/100.00))−1;
}
// the next function is needed for Turbo C++

// and Borland C++ to link in the appropriate
// functions for fscanf floating point formats:
static void force_fpf()
{
float x, *y;
y=&x;
x=*y;
}
// −−−−−−−−−−−−−−−−−−−−
// input layer
//−−−−−−−−−−−−−−−−−−−−−
input_layer::input_layer(int i, int o)
{
num_inputs=i;
num_outputs=o;
outputs = new float[num_outputs];

if (outputs==0)
{
cout << "not enough memory\n";
cout << "choose a smaller architecture\n";
exit(1);
}
}
input_layer::~input_layer()
{
delete [num_outputs] outputs;
}
void input_layer::calc_out()
{
//nothing to do, yet
}
// −−−−−−−−−−−−−−−−−−−−
// output layer
//−−−−−−−−−−−−−−−−−−−−−
output_layer::output_layer(int i, int o)
{
num_inputs =i;
num_outputs =o;
weights = new float[num_inputs*num_outputs];
output_errors = new float[num_outputs];
back_errors = new float[num_inputs];
expected_values = new float[num_outputs];
if ((weights==0)||(output_errors==0)||(back_errors==0)
||(outputs==0)||(expected_values==0))
{

exit(1);
}
}
output_layer::~output_layer()
{
// some compilers may require the array
// size in the delete statement; those
// conforming to Ansi C++ will not
delete [num_outputs*num_inputs] weights;
delete [num_outputs] output_errors;
delete [num_inputs] back_errors;
void output_layer::calc_out()
{
int i,j,k;
float accumulator=0.0;
for (j=0; j<num_outputs; j++)

{
for (i=0; i<num_inputs; i++)
{
k=i*num_outputs;
if (weights[k+j]*weights[k+j] > 1000000.0)
{
cout << "weights are blowing up\n";
cout << "try a smaller learning constant\n";
cout << "e.g. beta=0.02 aborting...\n";
exit(1);
}
outputs[j]=weights[k+j]*(*(inputs+i));
accumulator+=outputs[j];
}
// use the sigmoid squash function

outputs[j]=squash(accumulator);
accumulator=0;
}
}
void output_layer::calc_error(float & error)

{
int i, j, k;
float accumulator=0;
float total_error=0;

{
output_errors[j] = expected_values[j]−outputs[j];
total_error+=output_errors[j];
}
error=total_error;

{

k=i*num_outputs;
{
back_errors[i]=
weights[k+j]*output_errors[j];
accumulator+=back_errors[i];
}
back_errors[i]=accumulator;
accumulator=0;
// now multiply by derivative of
// sigmoid squashing function, which is
// just the input*(1−input)
back_errors[i]*=(*(inputs+i))*(1−(*(inputs+i)));
}
void output_layer::randomize_weights()
{
int i, j, k;
const unsigned first_time=1;
const unsigned not_first_time=0;

float discard;
discard=randomweight(first_time);
for (i=0; i< num_inputs; i++)

{
k=i*num_outputs;
for (j=0; j< num_outputs; j++)
weights[k+j]=randomweight(not_first_time);
}
}
void output_layer::update_weights(const float beta)

{
int i, j, k;
// learning law: weight_change =

// beta*output_error*input

{
k=i*num_outputs;
weights[k+j] +=
beta*output_errors[i]*(*(inputs+i));
}
void output_layer::list_weights()
{
int i, j, k;

{
k=i*num_outputs;
cout << "weight["<<i<<","<<
j<<"] is: "<<weights[k+j];

}
}
void output_layer::list_errors()
{
int i, j;

cout << "backerror["<<i<<
"] is : "<<back_errors[i]<<"\n";

cout << "outputerrors["<<j<<
"] is: "<<output_errors[j]<<"\n";
}
void output_layer::write_weights(int layer_no,

FILE * weights_file_ptr)
{
int i, j, k;
// assume file is already open and ready for

// writing
// prepend the layer_no to all lines of data

// format:
// layer_no weight[0,0] weight[0,1] ...
// ...

{
fprintf(weights_file_ptr,"%i ",layer_no);
k=i*num_outputs;
{
fprintf(weights_file_ptr,"%f",
weights[k+j]);
}
fprintf(weights_file_ptr,"\n");
}
}
void output_layer::read_weights(int layer_no,

{
int i, j, k;

// reading
// look for the prepended layer_no

// format:
// ...
while (1)
fscanf(weights_file_ptr,"%i",&j);
if ((j==layer_no)|| (feof(weights_file_ptr)))

break;
else
{
while (fgetc(weights_file_ptr) != `\n')
{;}// get rest of line
}
}
if (!(feof(weights_file_ptr)))
{
// continue getting first line
i=0;
{
fscanf(weights_file_ptr,"%f",
&weights[j]); // i*num_outputs = 0
}
fscanf(weights_file_ptr,"\n");
// now get the other lines

{
fscanf(weights_file_ptr,"%i",&layer_no);
k=i*num_outputs;
{
&weights[k+j]);
}
}
}
else cout << "end of file reached\n";
}
void output_layer::list_outputs()
{
int j;

{
cout << "outputs["<<j
<<"] is: "<<outputs[j]<<"\n";
}
// −−−−−−−−−−−−−−−−−−−−−
// middle layer
//−−−−−−−−−−−−−−−−−−−−−−
middle_layer::middle_layer(int i, int o):
output_layer(i,o)
{
middle_layer::~middle_layer()
{


}
void middle_layer::calc_error()
{
int i, j, k;

{
k=i*num_outputs;
{
back_errors[i]=
weights[k+j]*(*(output_errors+j));
}
accumulator=0;
}
}
network::network()
{
position=0L;
}
network::~network()
{
int i,j,k;
i=layer_ptr[0]−>num_outputs;// inputs
j=layer_ptr[number_of_layers−1]−>num_outputs; //outputs
k=MAX_VECTORS;
delete [(i+j)*k]buffer;
}
void network::set_training(const unsigned & value)

{
training=value;
}
unsigned network::get_training_value()
{
return training;
}
void network::get_layer_info()
{
int i;
//−−−−−−−−−−−−−−−−−−−−−
//
// Get layer sizes for the network
//

// −−−−−−−−−−−−−−−−−−−−
cout << " Please enter in the number of layers for your network.\n";
cout << " You can have a minimum of 3 to a maximum of 5. \n";
cout << " 3 implies 1 hidden layer; 5 implies 3 hidden layers : \n\n";
cin >> number_of_layers;
cout << " Enter in the layer sizes separated by spaces.\n";

cout << " For a network with 3 neurons in the input layer,\n";
cout << " 2 neurons in a hidden layer, and 4 neurons in the\n";
cout << " output layer, you would enter: 3 2 4 .\n";
cout << " You can have up to 3 hidden layers,for five maximum entries
:\n\n";
for (i=0; i<number_of_layers; i++)

{
cin >> layer_size[i];
}
// −−−−−−−−−−−−−−−−−−−−−−−−−−
// size of layers:
// input_layer layer_size[0]
// output_layer layer_size[number_of_layers−1]
// middle_layers layer_size[1]
// optional: layer_size[number_of_layers−3]
//−−−−−−−−−−−−−−−−−−−−−−−−−−−
}
void network::set_up_network()
{
int i,j,k;
//−−−−−−−−−−−−−−−−−−−−−−−−−−−
// Construct the layers
//
//−−−−−−−−−−−−−−−−−−−−−−−−−−−
layer_ptr[0] = new input_layer(0,layer_size[0]);
for (i=0;i<(number_of_layers−1);i++)
{
layer_ptr[i+1y] =
new middle_layer(layer_size[i],layer_size[i+1]);
}
layer_ptr[number_of_layers−1] = new
output_layer(layer_size[number_of_layers−2],layer_size[number_of_layers−
1]);
{
if (layer_ptr[i] == 0)
{
cout << "insufficient memory\n";
cout << "use a smaller architecture\n";
exit(1);
}
}
//−−−−−−−−−−−−−−−−−−−−−−−−−−
// Connect the layers

//
//−−−−−−−−−−−−−−−−−−−−−−−−−−
// set inputs to previous layer outputs for all layers,
// except the input layer
for (i=1; i< number_of_layers; i++)

layer_ptr[i]−>inputs = layer_ptr[i−1y]−>outputs;
// for back_propagation, set output_errors to next layer

// back_errors for all layers except the output
// layer and input layer
for (i=1; i< number_of_layers −1; i++)

((output_layer *)layer_ptr[i])−>output_errors =
((output_layer *)layer_ptr[i+1])−>back_errors;
// define the IObuffer that caches data from

// the datafile
k=MAX_VECTORS;
buffer=new
float[(i+j)*k];
if (buffer==0)
cout << "insufficient memory for buffer\n";
}
void network::randomize_weights()
{
int i;

((output_layer *)layer_ptr[i])
−>randomize_weights();
}
void network::update_weights(const float beta)

{
int i;

−>update_weights(beta);
}
void network::write_weights(FILE * weights_file_ptr)

{
int i;

−>write_weights(i,weights_file_ptr);
}
void network::read_weights(FILE * weights_file_ptr)

{
int i;

−>read_weights(i,weights_file_ptr);

void network::list_weights()
{
int i;

{
cout << "layer number : " <<i<< "\n";
−>list_weights();
}
}
void network::list_outputs()
{
int i;

{
−>list_outputs();
}
}
void network::write_outputs(FILE *outfile)

{
int i, ins, outs;
ins=layer_ptr[0]−>num_outputs;
outs=layer_ptr[number_of_layers−1]−>num_outputs;
float temp;
fprintf(outfile,"for input vector:\n");
for (i=0; i<ins; i++)

{
temp=layer_ptr[0]−>outputs[i];
fprintf(outfile,"%f ",temp);
}
fprintf(outfile,"\noutput vector is:\n");
for (i=0; i<outs; i++)

{
temp=layer_ptr[number_of_layers−1]−>
outputs[i];
if (training==1)
{
fprintf(outfile,"\nexpected output vector is:\n");

{
temp=((output_layer *)(layer_ptr[number_of_layers−1]))−>
expected_values[i];

fprintf(outfile,"\n−−−−−−−−−−\n");
void network::list_errors()
{
int i;

{
−>list_errors();
}
}
int network::fill_IObuffer(FILE * inputfile)

{
// this routine fills memory with
// an array of input, output vectors
// up to a maximum capacity of
// MAX_INPUT_VECTORS_IN_ARRAY
// the return value is the number of read
// vectors
int i, k, count, veclength;
int ins, outs;
if (training==1)
veclength=ins+outs;
else
veclength=ins;
count=0;
while ((count<MAX_VECTORS)&&
(!feof(inputfile)))
{
k=count*(veclength);
for (i=0; i<veclength; i++)
{
fscanf(inputfile,"%f",&buffer[k+i]);
}
fscanf(inputfile,"\n");
count++;
}
if (!(ferror(inputfile)))
return count;
else return −1; // error condition
void network::set_up_pattern(int buffer_index)

{

// read one vector into the network

int i, k;
int ins, outs;
if (training==1)
k=buffer_index*(ins+outs);
else
k=buffer_index*ins;

layer_ptr[0]−>outputs[i]=buffer[k+i];
if (training==1)
{
((output_layer *)layer_ptr[number_of_layers−1])−>
expected_values[i]=buffer[k+i+ins];
}
void network::forward_prop()
{
int i;
{
layer_ptr[i]−>calc_out(); //polymorphic
// function
}
}
void network::backward_prop(float & toterror)

{
int i;
// error for the output layer

((output_layer*)layer_ptr[number_of_layers−1])−>
calc_error(toterror);
// error for the middle layer(s)

for (i=number_of_layers−2; i>0; i−−)
{
((middle_layer*)layer_ptr[i])−>
calc_error();
}


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
A Look at the Functions in the layer.cpp File

The following is a listing of the functions in the layer.cpp file along with a brief statement of each one's
purpose.
• void set_training(const unsigned &) Sets the value of the private data member, training; use 1
for training mode, and 0 for test mode.
• unsigned get_training_value() Gets the value of the training constant that gives the mode in use.
• void get_layer_info() Gets information about the number of layers and layer sizes from the user.
• void set_up_network() This routine sets up the connections between layers by assigning pointers
appropriately.
• void randomize_weights() At the beginning of the training process, this routine is used to
randomize all of the weights in the network.
• void update_weights(const float) As part of training, weights are updated according to the
learning law used in backpropagation.
• void write_weights(FILE *) This routine is used to write weights to a file.
• void read_weights(FILE *) This routine is used to read weights into the network from a file.
• void list_weights() This routine can be used to list weights while a simulation is in progress.
• void write_outputs(FILE *) This routine writes the outputs of the network to a file.
• void list_outputs() This routine can be used to list the outputs of the network while a simulation
is in progress.
• void list_errors() Lists errors for all layers while a simulation is in progress.
• void forward_prop() Performs the forward propagation.
• void backward_prop(float &) Performs the backward error propagation.
• int fill_IObuffer(FILE *) This routine fills the internal IO buffer with data from the training or
test data sets.
• void set_up_pattern(int) This routine is used to set up one pattern from the IO buffer for
training.
• inline float squash(float input) This function performs the sigmoid function.
• inline float randomweight (unsigned unit) This routine returns a random weight between –1 and
1; use 1 to initialize the generator, and 0 for all subsequent calls.
Note that the functions squash(float) and randomweight(unsigned) are declared inline. This
means that the function's source code is inserted wherever it appears. This increases code size,
but also increases speed because a function call, which is expensive, is avoided.
The final file to look at is the backprop.cpp file presented in Listing 7.3.
Listing 7.3 The backprop.cpp file for the backpropagation simulator
// backprop.cpp V. Rao, H. Rao

#include "layer.cpp"
#define TRAINING_FILE "training.dat"

#define WEIGHTS_FILE "weights.dat"
#define OUTPUT_FILE "output.dat"
#define TEST_FILE "test.dat"
void main()
{
float error_tolerance =0.1;

float total_error =0.0;
float avg_error_per_cycle =0.0;
float error_last_cycle =0.0;
float avgerr_per_pattern =0.0; // for the latest cycle
float error_last_pattern =0.0;
float learning_parameter =0.02;
unsigned temp, startup;
long int vectors_in_buffer;
long int max_cycles;
long int patterns_per_cycle =0;
long int total_cycles, total_patterns;

int i;
// create a network object

network backp;
FILE * training_file_ptr, * weights_file_ptr, * output_file_ptr;

FILE * test_file_ptr, * data_file_ptr;
// open output file for writing

if ((output_file_ptr=fopen(OUTPUT_FILE,"w"))==NULL)
{
cout << "problem opening output file\n";
exit(1);
}
// enter the training mode : 1=training on 0=training off

cout << "−−−−−−−−−−−−−−−−−−−−−−−−\n";
cout << " C++ Neural Networks and Fuzzy Logic \n";
cout << " Backpropagation simulator \n";
cout << " version 1 \n";
cout << "−−−−−−−−−−−−−−−−−−−−−−−−\n";
cout << "Please enter 1 for TRAINING on, or 0 for off: \n\n";
cout << "Use training to change weights according to your\n";
cout << "expected outputs. Your training.dat file should contain\n";
cout << "a set of inputs and expected outputs. The number of\n";
cout << "inputs determines the size of the first (input) layer\n";
cout << "while the number of outputs determines the size of the\n";
cout << "last (output) layer :\n\n";
cin >> temp;

backp.set_training(temp);
if (backp.get_training_value() == 1)
{
cout << "−−> Training mode is *ON*. weights will be saved\n";
cout << "in the file weights.dat at the end of the\n";
cout << "current set of input (training) data\n";
}
else

{
cout << "−−> Training mode is *OFF*. weights will be loaded\n";
cout << "from the file weights.dat and the current\n";
cout << "(test) data set will be used. For the test\n";
cout << "data set, the test.dat file should contain\n";
cout << "only inputs, and no expected outputs.\n";
}
if (backp.get_training_value()==1)
{
// −−−−−−−−−−−−−−−−−−−−
// Read in values for the error_tolerance,
// and the learning_parameter
// −−−−−−−−−−−−−−−−−−−−
cout << " Please enter in the error_tolerance\n";
cout << " −−− between 0.001 to 100.0, try 0.1 to start \n";
cout << "\n";
cout << "and the learning_parameter, beta\n";
cout << " −−− between 0.01 to 1.0, try 0.5 to start −− \n\n";
cout << " separate entries by a space\n";
cout << " example: 0.1 0.5 sets defaults mentioned :\n\n";
cin >> error_tolerance >> learning_parameter;

//−−−−−−−−−−−−−−−−−−−−−
// open training file for reading
//−−−−−−−−−−−−−−−−−−−−
if ((training_file_ptr=fopen(TRAINING_FILE,"r"))==NULL)
{
cout << "problem opening training file\n";
exit(1);
}
data_file_ptr=training_file_ptr; // training on
// Read in the maximum number of cycles

// each pass through the input data file is a cycle
cout << "Please enter the maximum cycles for the simula−\
tion\n";
cout << "A cycle is one pass through the data set.\n";
cout << "Try a value of 10 to start with\n";
cin >> max_cycles;
}
else
{
if ((test_file_ptr=fopen(TEST_FILE,"r"))==NULL)
{
cout << "problem opening test file\n";
exit(1);
}
data_file_ptr=test_file_ptr; // training off
}
//
// training: continue looping until the total error is less than
// the tolerance specified, or the maximum number of
// cycles is exceeded; use both the forward signal propaga
tion
// and the backward error propagation phases. If the error
// tolerance criteria is satisfied, save the weights in a
file.
// no training: just proceed through the input data set once in the

// forward signal propagation phase only. Read the starting

// weights from a file.
// in both cases report the outputs on the screen
// initialize counters
total_cycles=0; // a cycle is once through all the input data
total_patterns=0; // a pattern is one entry in the input data
// get layer information

backp.get_layer_info();
// set up the network connections

backp.set_up_network();
// initialize the weights

{
// randomize weights for all layers; there is no
// weight matrix associated with the input layer
// weight file will be written after processing
// so open for writing
if ((weights_file_ptr=fopen(WEIGHTS_FILE,"w"))
==NULL)
{
cout << "problem opening weights file\n";
exit(1);
}
backp.randomize_weights();
}
else
{
// read in the weight matrix defined by a
// prior run of the backpropagation simulator
// with training on
if ((weights_file_ptr=fopen(WEIGHTS_FILE,"r"))
==NULL)
{
cout << "problem opening weights file\n";
exit(1);
}
backp.read_weights(weights_file_ptr);
}
// main loop
// if training is on, keep going through the input data
// until the error is acceptable or the maximum number of
// cycles
// is exceeded.
// if training is off, go through the input data once. report // outputs
// with inputs to file output.dat
startup=1;
vectors_in_buffer = MAX_VECTORS; // startup condition
total_error = 0;
while ( ((backp.get_training_value()==1)
&& (avgerr_per_pattern
> error_tolerance)
&& (total_cycles < max_cycles)
&& (vectors_in_buffer !=0))

|| ((backp.get_training_value()==0)
&& (total_cycles < 1))
&& (startup==1))
)
{
startup=0;
error_last_cycle=0; // reset for each cycle
patterns_per_cycle=0;
// process all the vectors in the datafile
// going through one buffer at a time
// pattern by pattern
while ((vectors_in_buffer==MAX_VECTORS))
{
vectors_in_buffer=
backp.fill_IObuffer(data_file_ptr); // fill buffer
if (vectors_in_buffer < 0)
{
cout << "error in reading in vectors, aborting\n";
cout << "check that there are no extra
linefeeds\n";
cout << "in your data file, and that the
number\n";
cout << "of layers and size of layers match
the\n";
cout << "the parameters provided.\n";
exit(1);
}
// process vectors
for (i=0; i<vectors_in_buffer; i++)
{
// get next pattern
backp.set_up_pattern(i);
total_patterns++;
patterns_per_cycle++;
// forward propagate
backp.forward_prop();
backp.write_outputs(output_file_ptr);
// back_propagate, if appropriate
{
backp.backward_prop(error_last_pattern);
error_last_cycle += error_last_pattern
z *error_last_pattern;
backp.update_weights(learning_parameter);
// backp.list_weights();
// can
// see change in weights by
// using list_weights before and
// after back_propagation
}

error_last_pattern = 0;
}
avgerr_per_pattern=((float)sqrt((double)error_last_cycle
/patterns_per_cycle));
total_error += error_last_cycle;
total_cycles++;
// most character displays are 26 lines

// user will see a corner display of the cycle count
// as it changes
cout << "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n";

cout << total_cycles << "\t" << avgerr_per_pattern << "\n";
fseek(data_file_ptr, 0L, SEEK_SET); // reset the file pointer

// to the beginning of
// the file
vectors_in_buffer = MAX_VECTORS; // reset
} // end main loop
cout << "\n\n\n\n\n\n\n\n\n\n\n";

cout << "−−−−−−−−−−−−−−−−−−−−−−−−\n";
cout << " done: results in file output.dat\n";
cout << " training: last vector only\n";
cout << " not training: full cycle\n\n";
{
backp.write_weights(weights_file_ptr);
avg_error_per_cycle = (float)sqrt((double)total_error/
total_cycles);
error_last_cycle = (float)sqrt((double)error_last_cycle);
cout << " weights saved in file weights.dat\n";

cout << "\n";
cout << "−−>average error per cycle = " <<
avg_error_per_cycle << " <−\n";
cout << "−−>error last cycle= " << error_last_cycle << " <−\n";
cout << "−>error last cycle per pattern= " << avgerr_per_pattern << " <−
\n";
cout << "−−−−−−>total cycles = " << total_cycles << " <−−\n";
cout << "−−−−−−>total patterns = " << total_patterns << " <−−−\n";
cout << "−−−−−−−−−−−−−−−−−−−−−−−−−\n";
// close all files
fclose(data_file_ptr);
fclose(weights_file_ptr);
fclose(output_file_ptr);


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
The backprop.cpp file implements the simulator controls. First, data is accepted from the user for network
parameters. Assuming Training mode is used, the training file is opened and data is read from the file to fill
the IO buffer. Then the main loop is executed where the network processes pattern by pattern to complete a
cycle, which is one pass through the entire training data set. (The IO buffer is refilled as required during this
process.) After executing one cycle, the file pointer is reset to the beginning of the file and another cycle
begins. The simulator continues with cycles until one of the two fundamental criteria is met:
1. The maximum cycle count specified by the user is exceeded.

2. The average error per pattern for the latest cycle is less than the error tolerance specified by the
user.
When either of these occurs, the simulator stops and reports out the error achieved, and saves weights in the
weights.dat file and one output vector in the output.dat file.
In Test mode, exactly one cycle is processed by the network and outputs are written to the output.dat file. At
the beginning of the simulation in Test mode, the network is set up with weights from the weights.dat file. To
simplify the program, the user is requested to enter the number of layers and size of layers, although you
could have the program figure this out from the weights file.
Compiling and Running the Backpropagation Simulator

Compiling the backprop.cpp file will compile the simulator since layer.cpp is included in backprop.cpp. To
run the simulator, once you have created an executable (using 80X87 floating point hardware if available),
you type in backprop and see the following screen (user input in italic):

Backpropagation simulator
version 1
Please enter 1 for TRAINING on, or 0 for off:
Use training to change weights according to your

expected outputs. Your training.dat file should contain
a set of inputs and expected outputs. The number of
inputs determines the size of the first (input) layer
while the number of outputs determines the size of the
last (output) layer :
−> Training mode is *ON*. weights will be saved

in the file weights.dat at the end of the
current set of input (training) data
Please enter in the error_tolerance
−− between 0.001 to 100.0, try 0.1 to start −−
and the learning_parameter, beta

−− between 0.01 to 1.0, try 0.5 to start −−
separate entries by a space

example: 0.1 0.5 sets defaults mentioned :
0.2 0.25
Please enter the maximum cycles for the simulation

A cycle is one pass through the data set.
Try a value of 10 to start with
Please enter in the number of layers for your network.
You can have a minimum of three to a maximum of five.
three implies one hidden layer; five implies three hidden layers:
Enter in the layer sizes separated by spaces.

For a network with three neurons in the input layer,
two neurons in a hidden layer, and four neurons in the
output layer, you would enter: 3 2 4.
You can have up to three hidden layers for five maximum entries :
2 2 1
1 0.353248
2 0.352684
3 0.352113
4 0.351536
5 0.350954
...
299 0.0582381
300 0.0577085
−−−−−−−−−−−−−−−−−−−−−−−−
done: results in file output.dat
training: last vector only
not training: full cycle
weights saved in file weights.dat

−−>average error per cycle = 0.20268 <−−
−−>error last cycle = 0.0577085 <−−
−>error last cycle per pattern= 0.0577085 <−−
−−−−−−>total cycles = 300 <−−
−−−−−−>total patterns = 300 <−−
The cycle number and the average error per pattern is displayed as the simulation progresses
(not all values shown). You can monitor this to make sure the simulator is converging on a
solution. If the error does not seem to decrease beyond a certain point, but instead drifts or
blows up, then you should start the simulator again with a new starting point defined by the
random weights initializer. Also, you could try decreasing the size of the learning rate
parameter. Learning may be slower, but this may allow a better minimum to be found.
This example shows just one pattern in the training set with two inputs and one output. The results along with
the (one) last pattern are shown as follows from the file output.dat:
for input vector:

0.400000 −0.400000
output vector is:
0.842291


0.900000
The match is pretty good, as can be expected, since the optimization is easy for the network; there is only one
pattern to worry about. Let’s look at the final set of weights for this simulation in weights.dat. These weights
were obtained by updating the weights for 300 cycles with the learning law:
1 0.175039 0.435039
1 −1.319244 −0.559244
2 0.358281
2 2.421172
We’ll leave the backpropagation simulator for now and return to it in a later chapter for further exploration.
You can experiment a number of different ways with the simulator:
• Try a different number of layers and layer sizes for a given problem.
• Try different learning rate parameters and see its effect on convergence and training time.
• Try a very large learning rate parameter (should be between 0 and 1); try a number over 1 and
note the result.


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Summary
In this chapter, you learned about one of the most powerful neural network algorithms called
backpropagation. Without having feedback connections, propagating only errors appropriately to the hidden
layer and input layer connections, the algorithm uses the so−called generalized delta rule and trains the
network with exemplar pairs of patterns. It is difficult to determine how many hidden−layer neurons are to be
provided for. The number of hidden layers could be more than one. In general, the size of the hidden layer(s)
is related to the features or distinguishing characteristics that should be discerned from the data. Our example
in this chapter relates to a simple case where there is a single hidden layer. The outputs of the output neurons,
and therefore of the network, are vectors with components between 0 and 1, since the thresholding function
is the sigmoid function. These values can be scaled, if necessary, to get values in another interval.
Our example does not relate to any particular function to be computed by the network, but inputs and outputs
were randomly chosen. What this can tell you is that, if you do not know the functional equation between two
sets of vectors, the feedback backpropagation network can find the mapping for any vector in the domain,
even if the functional equation is not found. For all we know, that function could be nonlinear as well.
There is one important fact you need to remember about the backpropagation algorithm. Its steepest descent
procedure in training does not guarantee finding a global or overall minimum, it can find only a local
minimum of the energy surface.
Summary 151

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Chapter 8
BAM: Bidirectional Associative Memory
Introduction
The Bidirectional Associative Memory (BAM) model has a neural network of two layers and is fully connected
from each layer to the other. That is, there are feedback connections from the output layer to the input layer.
However, the weights on the connections between any two given neurons from different layers are the same.
You may even consider it to be a single bidirectional connection with a single weight. The matrix of weights
for the connections from the output layer to the input layer is simply the transpose of the matrix of weights for
the connections between the input and output layer. If we denote the matrix for forward connection weights by
W, then WT is the matrix of weights for the output layer to input layer connections. As you recall, the
transpose of a matrix is obtained simply by interchanging the rows and the columns of the matrix.
There are two layers of neurons, an input layer and an output layer. There are no lateral connections, that is,
no two neurons within the same layer are connected. Recurrent connections, which are feedback connections
to a neuron from itself, may or may not be present. The architecture is quite simple. Figure 8.1 shows the
layout for this neural network model, using only three input neurons and two output neurons. There are
feedback connections from Field A to Field B and vice−versa. This figure also indicates the presence of inputs
and outputs at each of the two fields for the bidirectional associative memory network. Connection weights
are also shown as labels on only a few connections in this figure, to avoid cluttering. The general case is
analogous.
Figure 8.1 Layout of a BAM network
Inputs and Outputs

The input to a BAM network is a vector of real numbers, usually in the set { –1, +1 }. The output is also a
vector of real numbers, usually in the set { –1, +1 }, with the same or different dimension from that of the
input. These vectors can be considered patterns, and the network makes heteroassociation of patterns. If the
output is required to be the same as input, then you are asking the network to make autoassociation, which it
does, and it becomes a special case of the general activity of this type of neural network.
For inputs and outputs that are not outside the set containing just –1 and +1, try following this next procedure.
You can first make a mapping into binary numbers, and then a mapping of each binary digit into a bipolar
digit. For example, if your inputs are first names of people, each character in the name can be replaced by its
ASCII code, which in turn can be changed to a binary number, and then each binary digit 0 can be replaced by
Chapter 8 BAM: Bidirectional Associative Memory 152

–1. For example, the ASCII code for the letter R is 82, which is 1010010, as a binary number. This is mapped
onto the bipolar string 1 –1 1 –1 –1 1 –1. If a name consists of three characters, their ASCII codes in binary
can be concatenated or juxtaposed and the corresponding bipolar string obtained. This bipolar string can also
be looked upon as a vector of bipolar characters.
Weights and Training

BAM does not modify weights during its operation, and as mentioned in Chapter 6, like the Hopfield network,
uses one−shot training. The adaptive variety of BAM, called the Adaptive Bidirectional Associative Memory,
(ABAM) undergoes supervised iterative training. BAM needs some exemplar pairs of vectors. The pairs used
as exemplars are those that require heteroassociation. The weight matrix, there are two, but one is just the
transpose of the other as already mentioned, is constructed in terms of the exemplar vector pairs.
The use of exemplar vectors is a one−shot learning—to determine what the weights should be. Once weights
are so determined, and an input vector is presented, a potentially associated vector is output. It is taken as
input in the opposite direction, and its potentially associated vector is obtained back at the input layer. If the
last vector found is the same as what is originally input, then there is resonance. Suppose the vector B is
obtained at one end, as a result of C being input at the other end. If B in turn is input during the next cycle of
operation at the end where it was obtained, and produces C at the opposite end, then you have a pair of
heteroassociated vectors. This is what is basically happening in a BAM neural network.
NOTE: The BAM and Hopfield memories are closely related. You can think of the
Hopfield memory as a special case of the BAM.
What follow are the equations for the determination of the weight matrix, when the k pairs of exemplar
vectors are denoted by ( Xi, Yi), i ranging from 1 to k. Note that T in the superscript of a matrix stands for the
transpose of the matrix. While you interchange the rows and columns to get the transpose of a matrix, you
write a column vector as a row vector, and vice versa to get the transpose of a vector. The following equations
refer to the vector pairs after their components are changed to bipolar values, only for obtaining the weight
matrix W. Once W is obtained, further use of these exemplar vectors is made in their original form.
W = X1T Y1 + ... + XkT Yk
and
WT = Y1TX1 + ... + YkT Xk
Example
Suppose you choose two pairs of vectors as possible exemplars. Let them be:
X1 = (1, 0, 0, 1), Y1= (0, 1, 1)
and
X2 = (0, 1, 1, 0), Y2 = (1, 0, 1)
These you change into bipolar components and get, respectively, (1, –1, –1, 1), (–1, 1, 1), (–1, 1, 1, –1), and
(1, –1, 1).
Weights and Training 153

1 [−1 1 1] −1 [1 −1 1] −1 1 1 −1 1 −1 −2 2 0
W = −1 + 1 = 1 −1 −1 + 1 −1 1 = 2 −2 0
−1 1 1 −1 −1 1 −1 1 2 −2 0
1 −1 −1 1 1 −1 1 −1 −2 2 0
and
−2 2 2 −2
WT = 2 −2 −2 2
0 0 0 0


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
You may think that the last column of W being all zeros presents a problem in that when the input is X1 and
whatever the output, when this output is presented back in the backward direction, it does not produce X1,
which does not have a zero in the last component. There is a thresholding function needed here. The
thresholding function is transparent when the activations are all either +1 or –1. It then just looks like you do
an inverse mapping from bipolar values to binary, which is done by replacing each –1 by a 0. When the
activation is zero, you simply leave the output of that neuron as it was in the previous cycle or iteration. We
now present the thresholding function for BAM outputs.
1 if yj > 0 1 if xi > 0
bj|t+1 = bj|t if yj = 0 and ai|t+1 = ai|t if xi = 0
0 if yj < 0 0 if xi < 0
where xi and yj are the activations of neurons i and j in the input layer and output layer, respectively, and bj|t
refers to the output of the jth neuron in the output layer in the cycle t, while ai|t refers to the output of the ith
neuron in the input layer in the cycle t. Note that at the start, the ai and bj values are the same as the
corresponding components in the exemplar pair being used.
If X1 = (1, 0, 0, 1) is presented to the input neurons, their activations are given by the vector (–4, 4, 0). The
output vector, after using the threshold function just described is (0, 1, 1). The last component here is
supposed to be the same as the output in the previous cycle, since the corresponding activation value is 0.
Since X1 and Y1 are one exemplar pair, the third component of Y1 is what we need as the third component of
the output in the current cycle of operation; therefore, we fed X1 and received Y1. If we feed Y1 at the other
end (B field) , the activations in the A field will be (2, –2, –2, 2), and the output vector will be (1, 0, 0, 1),
which is X1.
With A field input X2 you get B field activations (4, −4, 0), giving the output vector as (1, 0, 1), which is Y2.
Thus X2 and Y2 are heteroassociated with each other.
Let us modify our X1 to be (1, 0, 1, 1). Then the weight matrix W becomes
1 [−1 1 1] −1 [1 −1 1] −2 2 0
W = −1 + 1 = 2 −2 0
1 1 0 0 2
1 −1 −2 2 0
and
−2 2 0 −2
WT = 2 −2 0 2
0 0 2 0
Now this is a different set of two exemplar vector pairs. The pairs are X1 = (1, 0, 1, 1), 1 = (0, 1, 1), and X2 =
(0, 1, 1, 0), Y2 = (1, 0, 1). Naturally, the weight matrix is different, as is its transpose, correspondingly. As
stated before, the weights do not change during the operation of the network with whatever inputs presented to

it. The results are as shown in Table 8.1 below.
Table 8.1 Results for the Example

Input vectoractivationoutput vector
X1 = (1, 0, 1, 1)(−4, 4, 2)(0, 1, 1) = Y1
X2 = (0, 1, 1, 0)(2, −2, 2)(1, 0, 1) = Y2
Y1 = (0, 1, 1)(2, −2, 2, 2)(1, 0, 1, 1) = X1
Y2 = (1, 0, 1)(−2, 2, 2, −2)(0, 1, 1,0) = X2
You may think that you will encounter a problem when you input a new vector and one of the neurons has
activation 0. In the original example, you did find this situation when you got the third output neuron’s
activation as 0. The thresholding function asked you to use the same output for this neuron as existed in the
earlier time cycle. So you took it to be 1, the third component in (0, 1, 1). But if your input vector is a new X
vector for which you are trying to find an associated Y vector, then you do not have a Y component to fall
back on when the activation turns out to be 0. How then can you use the thresholding function as stated?
What guidance do you have in this situation? If you keep track of the inputs used and outputs received thus
far, you realize that the Field B (where you get your Y vector) neurons are in some state, meaning that they
had some outputs perhaps with some training vector. If you use that output component as the one existing in
the previous cycle, you have no problem in using the thresholding function.
As an example, consider the input vector X3 = ( 1, 0, 0, 0), with which the activations of neurons in Field B
would be (−2, 2, 0). The first component of the output vector is clearly 0, and the second clearly 1. The third
component is what is in doubt. Considering the last row of the table where Y2 gives the state of the neurons in
Field B, you can accept 1, the last component of Y2, as the value you get from the thresholding function
corresponding to the activation value 0. So the output would be the vector (0, 1, 1), which is Y1. But Y1 is
heteroassociated with X1. Well, it means that X3 = ( 1, 0, 0, 0) is not heteroassociated with any X vector.


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Recall of Vectors
When X1 is presented at the input layer, the activation at the output layer will give ( –4, 4, 2) to which we
apply the thresholding function, which replaces a positive value by 1, and a negative value by 0.
This then gives us the vector (0, 1, 1) as the output, which is the same as our Y1. Now Y1 is passed back to the
input layer through the feedback connections, and the activation of the input layer becomes the vector (2, –2,
2, 2), which after thresholding gives the output vector (1, 0, 1, 1), same as X1. When X2 is presented at the
input layer, the activation at the output layer will give (2, –2, 2) to which the thresholding function, which
replaces a positive value by 1 and a negative value by 0, is applied. This then gives the vector (1, 0, 1) as the
output, which is the same as Y2. Now Y2 is passed back to the input layer through the feedback connections to
get the activation of the input layer as (–2, 2, 2, –2), which after thresholding gives the output vector (0, 1, 1,
0), which is X2.
The two vector pairs chosen here for encoding worked out fine, and the BAM network with
four neurons in Field A and three neurons in Field B is all set for finding a vector under
heteroassociation with a given input vector.
Continuation of Example
Let us now use the vector X3 = (1, 1, 1, 1). The vector Y3 = (0, 1, 1) is obtained at the output layer. But the
next step in which we present Y3 in the backward direction does not produce X3, instead it gives an X1 = (1, 0,
1, 1). We already have X1 associated with Y1. This means that X3 is not associated with any vector in the
output space. On the other hand, if instead of getting X1 we obtained a different X4 vector, and if this in the
feed forward operation produced a different Y vector, then we repeat the operation of the network until no
changes occur at either end. Then we will have possibly a new pair of vectors under the heteroassociation
established by this BAM network.
Special Case—Complements
If a pair of (distinct) patterns X and Y are found to be heteroassociated by BAM, and if you input the
complement of X, complement being obtained by interchanging the 0’s and 1’s in X, BAM will show that the
complement of Y is the pattern associated with the complement of X. An example will be seen in the
illustrative run of the program for C++ implementation of BAM, which follows.
C++ Implementation
In our C++ implementation of a discrete bidirectional associative memory network, we create classes for
neuron and network. Other classes created are called exemplar, assocpair, potlpair, for the exemplar pair
of vectors, associated pair of vectors, and potential pairs of vectors, respectively, for finding heteroassociation
between them. We could have made one class of pairvect for a pair of vectors and derived the exemplar and
Recall of Vectors 157

so on from it. The network class is declared as a friend class in these other classes. Now we present the
header and source files, called bamntwrk.h and bamntwrk.cpp. Since we reused our previous code from the
Hopfield network of Chapter 4, there are a few data members of classes that we did not put to explicit use in
the program. We call the neuron class bmneuron to remind us of BAM.
Program Details and Flow
A neuron in the first layer is referred to as anrn, and the number of neurons in this layer is referred to as
anmbr. We give the namebnrn to the array of neurons in the second layer, and bnmbr denotes the size of
that array. The sequence of operations in the program is as follows:
• We ask the user to input the exemplar vectors, and we transform them into their bipolar versions.
The trnsfrm ( ) function in the exemplar class is for this purpose.
• We give the network the X vector, in its bipolar version, in one exemplar pair. We find the
activations of the elements of bnrn array and get corresponding output vector as a binary pattern. If
this is the Y in the exemplar pair, the network has made a desired association in one direction, and we
go on to the next.step. Otherwise we have a potential associated pair, one of which is X and the other
is what we just got as the output vector in the opposite layer. We say potential associated pair because
we have the next step to confirm the association.
• We run the bnrn array through the transpose of the weight matrix and calculate the outputs of the
anrn array elements. If , as a result, we get the vector X as the anrn array, we found an associated
pair, (X, Y). Otherwise, we repeat the two steps just described until we find an associated pair.
• We now work with the next pair of exemplar vectors in the same manner as above, to find an
associated pair.
• We assign serial numbers, denoted by the variable idn, to the associated pairs so we can print
them all together at the end of the program. The pair is called (X, Y) where X produces Y through the
weight matrix W, and Y produces X through the weight matrix which is the transpose of W.
• A flag is used to have value 0 until confirmation of association is obtained, when the value of the
flag changes to 1.
• Functions compr1 and compr2 in the network class verify if the potential pair is indeed an
associated pair and set the proper value of the flag mentioned above.
• Functions comput1 and comput2 in the network class carry out the calculations to get the
activations and then find the output vector, in the proper directions of the bidirectional associative
memory network.
Program Details and Flow 158


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Program Example for BAM

For our illustration run, we provided for six neurons in the input layer and five in the output layer. We used
three pairs of exemplars for encoding. We used two additional input vectors, one of which is the complement
of the X of an exemplar pair, after the encoding is done, to see what association will be established in these
cases, or what recall will be made by the BAM network.
Header File
As expected, the complement of the Y of the exemplar is found to be associated with the complement of the X
of that pair. When the second input vector is presented, however, a new pair of associated vectors is found.
After the code is presented, we list the computer output also.
Listing 8.1 bamntwrk.h
//bamntwrk.h V. Rao, H. Rao
//Header file for BAM network program
#include <math.h>
#include <stdlib.h>
#define MXSIZ 10 // determines the maximum size of the network
class bmneuron
{
protected:
int nnbr;
int inn,outn;
int output;
int activation;
int outwt[MXSIZ];
char *name;
public:
bmneuron() { };
void getnrn(int,int,int,char *);
};
class exemplar
{
protected:
int xdim,ydim;
int v1[MXSIZ],v2[MXSIZ];
int u1[MXSIZ],u2[MXSIZ];
Program Example for BAM 159

friend class mtrx;
public:
exemplar() { };
void getexmplr(int,int,int *,int *);
void prexmplr();
void trnsfrm();
void prtrnsfrm();
};
class asscpair
{
protected:
int xdim,ydim,idn;
public:
asscpair() { };
void getasscpair(int,int,int);
void prasscpair();
};
class potlpair
{
protected:
int xdim,ydim;
public:
potlpair() { };
void getpotlpair(int,int);
void prpotlpair();
};
class network
{
public:
int anmbr,bnmbr,flag,nexmplr,nasspr,ninpt;
bmneuron (anrn)[MXSIZ],(bnrn)[MXSIZ];
exemplar (e)[MXSIZ];
asscpair (as)[MXSIZ];
potlpair (pp)[MXSIZ];
int outs1[MXSIZ],outs2[MXSIZ];
int mtrx1[MXSIZ][MXSIZ],mtrx2[MXSIZ][MXSIZ];
network() { };
void getnwk(int,int,int,int [][6],int [][5]);
void compr1(int,int);
void prwts();
void iterate();
void findassc(int *);
void asgninpt(int *);
void asgnvect(int,int *,int *);
void comput1();
void comput2();
void prstatus();
};
Program Example for BAM 160

Source File
The program source file is presented as follows.
Source File 161


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Listing 8.2 bamntwrk.cpp
//bamntwrk.cpp V. Rao, H. Rao
//Source file for BAM network program
#include “bamntwrk.h”
void bmneuron::getnrn(int m1,int m2,int m3,char *y)

{
int i;
name = y;
nnbr = m1;
outn = m2;
inn = m3;
for(i=0;i<outn;++i){
outwt[i] = 0 ;
}
output = 0;
activation = 0;
}
void exemplar::getexmplr(int k,int l,int *b1,int *b2)

{
int i2;
xdim = k;
ydim = l;
for(i2=0;i2<xdim;++i2){
v1[i2] = b1[i2]; }
for(i2=0;i2<ydim;++i2){
v2[i2] = b2[i2]; }
}
void exemplar::prexmplr()
{
int i;
cout<<”\nX vector you gave is:\n”;
for(i=0;i<xdim;++i){
cout<<v1[i]<<” “;}
cout<<”\nY vector you gave is:\n”;
for(i=0;i<ydim;++i){
cout<<v2[i]<<” “;}
cout<<”\n”;
}
Source File 162

void exemplar::trnsfrm()
{
int i;
u1[i] = 2*v1[i] −1;}
u2[i] = 2*v2[i] − 1;}
void exemplar::prtrnsfrm()
{
int i;
cout<<”\nbipolar version of X vector you gave is:\n”;
cout<<u1[i]<<” “;}
cout<<”\nbipolar version of Y vector you gave is:\n”;
cout<<u2[i]<<” “;}
cout<<”\n”;
}
void asscpair::getasscpair(int i,int j,int k)
{
idn = i;
xdim = j;
ydim = k;
}
void asscpair::prasscpair()
{
int i;
cout<<”\nX vector in the associated pair no. “<<idn<<” is:\n”;
cout<<v1[i]<<” “;}
cout<<”\nY vector in the associated pair no. “<<idn<<” is:\n”;
cout<<v2[i]<<” “;}
cout<<”\n”;
}
void potlpair::getpotlpair(int k,int j)

{
xdim = k;
ydim = j;
void potlpair::prpotlpair()
{
int i;
Source File 163

cout<<”\nX vector in possible associated pair is:\n”;
cout<<v1[i]<<” “;}
cout<<”\nY vector in possible associated pair is:\n”;
cout<<v2[i]<<” “;}
cout<<”\n”;
}
void network::getnwk(int k,int l,int k1,int b1[][6],int

b2[][5])
{
anmbr = k;
bnmbr = l;
nexmplr = k1;
nasspr = 0;
ninpt = 0;
int i,j,i2;
flag =0;
char *y1=”ANEURON”, *y2=”BNEURON” ;
for(i=0;i<nexmplr;++i){
e[i].getexmplr(anmbr,bnmbr,b1[i],b2[i]);
e[i].prexmplr();
e[i].trnsfrm();
e[i].prtrnsfrm();
}
for(i=0;i<anmbr;++i){
anrn[i].bmneuron::getnrn(i,bnmbr,0,y1);}
for(i=0;i<bnmbr;++i){
bnrn[i].bmneuron::getnrn(i,0,anmbr,y2);}
for(j=0;j<bnmbr;++j){
mtrx1[i][j] = 0;
for(i2=0;i2<nexmplr;++i2){
mtrx1[i][j] += e[i2].u1[i]*e[i2].u2[j];}
mtrx2[j][i] = mtrx1[i][j];
anrn[i].outwt[j] = mtrx1[i][j];
bnrn[j].outwt[i] = mtrx2[j][i];
}
}
prwts();
cout<<”\n”;
}
void network::asgninpt(int *b)

{
int i;
cout<<”\n”;
anrn[i].output = b[i];
outs1[i] = b[i];
Source File 164

void network::compr1(int j,int k)

{
int i;
if(pp[j].v1[i] != pp[k].v1[i]) flag = 1;
break;
}

{
int i;
break;}
void network::comput1()
{
int j;
int ii1;
int c1 =0,d1;
cout<<”\n”;
for(ii1=0;ii1<anmbr;++ii1){
d1 = outs1[ii1] * mtrx1[ii1][j];
c1 += d1;
}
bnrn[j].activation = c1;
cout<<”\n output layer neuron “<<j<<” activation is”
<<c1<<”\n”;
if(bnrn[j].activation <0) {
bnrn[j].output = 0;
outs2[j] = 0;}
else
if(bnrn[j].activation>0) {
bnrn[j].output = 1;
outs2[j] = 1;}
else
{cout<<”\n A 0 is obtained, use previous output

value \n”;
if(ninpt<=nexmplr){
bnrn[j].output = e[ninpt−1].v2[j];}
Source File 165

else
{ bnrn[j].output = pp[0].v2[j];}
outs2[j] = bnrn[j].output; }
cout<<”\n output layer neuron “<<j<<” output is”

<<bnrn[j].output<<”\n”;
}
}
void network::comput2()
{
int i;
int ii1;
int c1=0;
for(ii1=0;ii1<bnmbr;++ii1){
c1 += outs2[ii1] * mtrx2[ii1][i]; }
anrn[i].activation = c1;
cout<<”\ninput layer neuron “<<i<<”activation is “
<<c1<<”\n”;
if(anrn[i].activation <0 ){
anrn[i].output = 0;
outs1[i] = 0;}
else
if(anrn[i].activation >0 ) {
anrn[i].output = 1;
outs1[i] = 1;
}
else
{ cout<<”\n A 0 is obtained, use previous value if available\n”;
if(ninpt<=nexmplr){
anrn[i].output = e[ninpt−1].v1[i];}
else
{anrn[i].output = pp[0].v1[i];}
outs1[i] = anrn[i].output;}
cout<<”\n input layer neuron
“<<i<<” output is “
<<anrn[i].output<<”\n”;
}
}
void network::asgnvect(int j1,int *b1,int *b2)

{
int j2;
for(j2=0;j2<j1;++j2){
Source File 166

b2[j2] = b1[j2];}
void network::prwts()
{
int i3,i4;
cout<<”\n weights— input layer to output layer: \n\n”;
for(i3=0;i3<anmbr;++i3){
for(i4=0;i4<bnmbr;++i4){
cout<<anrn[i3].outwt[i4]<<” “;}
cout<<”\n”; }
cout<<”\n”;
cout<<”\nweights— output layer to input layer: \n\n”;
cout<<bnrn[i3].outwt[i4]<<” “;}
cout<<”\n”; }
cout<<”\n”;
}
void network::iterate()
{
int i1;
findassc(e[i1].v1);
}
void network::findassc(int *b)

{
int j;
flag = 0;
asgninpt(b);
ninpt ++;
cout<<”\nInput vector is:\n” ;
for(j=0;j<6;++j){
cout<<b[j]<<” “;}
cout<<”\n”;
pp[0].getpotlpair(anmbr,bnmbr);
asgnvect(anmbr,outs1,pp[0].v1);
comput1();
if(flag>=0){
asgnvect(bnmbr,outs2,pp[0].v2);
cout<<”\n”;
pp[0].prpotlpair();
Source File 167

cout<<”\n”;
comput2(); }
for(j=1;j<MXSIZ;++j){
pp[j].getpotlpair(anmbr,bnmbr);
asgnvect(anmbr,outs1,pp[j].v1);
comput1();
asgnvect(bnmbr,outs2,pp[j].v2);
pp[j].prpotlpair();
cout<<”\n”;
compr1(j,j−1);
compr2(j,j−1);
if(flag == 0) {
int j2;
nasspr += 1;
j2 = nasspr;
as[j2].getasscpair(j2,anmbr,bnmbr);
asgnvect(anmbr,pp[j].v1,as[j2].v1);
asgnvect(bnmbr,pp[j].v2,as[j2].v2);
cout<<”\nPATTERNS ASSOCIATED:\n”;
as[j2].prasscpair();
j = MXSIZ ;
}
else
if(flag == 1)
{
flag = 0;
comput1();
}
}
}
void network::prstatus()
{
int j;
cout<<”\nTHE FOLLOWING ASSOCIATED PAIRS WERE FOUND BY BAM\n\n”;
for(j=1;j<=nasspr;++j){
as[j].prasscpair();
cout<<”\n”;}
void main()
{
int ar = 6, br = 5, nex = 3;
int inptv[][6]={1,0,1,0,1,0,1,1,1,0,0,0,0,1,1,0,0,0,0,1,0,1,0,1,
1,1,1,1,1,1};
Source File 168

int outv[][5]={1,1,0,0,1,0,1,0,1,1,1,0,0,1,0};
cout<<”\n\nTHIS PROGRAM IS FOR A BIDIRECTIONAL ASSOCIATIVE MEMORY NETWORK.\n”;

cout<<” THE NETWORK ISSET UP FOR ILLUSTRATION WITH “<<ar<<
“ INPUT NEURONS, AND “<<br;
cout<<” OUTPUT NEURONS.\n”<<nex
<<” exemplars are used to encode \n”;
static network bamn;

bamn.getnwk(ar,br,nex,inptv,outv) ;
bamn.iterate();
bamn.findassc(inptv[3]);
bamn.findassc(inptv[4]);
bamn.prstatus();
}
Source File 169


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Program Output
The output from an illustrative run of the program is listed next. You will notice that we provided for a lot of
information to be output as the program is executed. The fourth vector we input is not from an exemplar, but it
is the complement of the X of the first exemplar pair. The network found the complement of the Y of this
exemplar to be associated with this input. The fifth vector we input is (1, 1, 1, 1, 1, 1). But BAM recalled in
this case the complement pair for the third exemplar pair, which is X = (1, 0, 0, 1, 1, 1) and Y = (0, 1, 1, 0, 1).
You notice that the Hamming distance of this input pattern from the X’s of the three exemplars are 3, 3, and 4,
respectively. The hamming distance from the complement of the X of the first exemplar pair is also 3. But the
Hamming distance of (1, 1, 1, 1) from the X of the complement of the X of the third exemplar pair is only 2. It
would be instructive if you would use the input pattern (1, 1, 1, 1, 1, 0) to see to what associated pair it leads.
Now the output:
THIS PROGRAM IS FOR A BIDIRECTIONAL ASSOCIATIVE

MEMORY NETWORK.
THE NETWORK IS SET UP FOR ILLUSTRATION WITH SIX INPUT
NEURONS AND FIVE OUTPUT NEURONS.
Three exemplars are used to encode
X vector you gave is:

1 0 1 0 1 0
Y vector you gave is:
1 1 0 0 1
bipolar version of X vector you gave is:

1 −1 1 −1 1 −1
bipolar version of Y vector you gave is:
1 1 −1 −1 1

1 1 1 0 0 0
0 1 0 1 1

1 1 1 −1 −1 −1
−1 1 −1 1 1

0 1 1 0 0 0
1 0 0 1 0

−1 1 1 −1 −1 −1
1 −1 −1 1 −1
Program Output 170

weights— input layer to output layer:
−1 3 −1 −1 3
−1 1 −1 3 −1
1 1 −3 1 1
−1 −1 3 −1 −1
1 1 1 −3 1
−1 −1 3 −1 −1
weights— output layer to input layer:
−1 −1 1 −1 1 −1
3 −1 1 −1 1 −1
−1 −1 −3 3 1 3
−1 3 1 −1 −3 −1
3 −1 1 −1 1 −1
Input vector is:

1 0 1 0 1 0
output layer neuron 0 activation is 1
output layer neuron 0 output is 1
output layer neuron 2 activation is −3
X vector in possible associated pair is:

1 0 1 0 1 0
Y vector in possible associated pair is:
1 1 0 0 1
input layer neuron 0 activation is 5
input layer neuron 0 output is 1
input layer neuron 1 activation is −3
Program Output 171


1 0 1 0 1 0
1 1 0 0 1
PATTERNS ASSOCIATED:
X vector in the associated pair no. 1 is:

1 0 1 0 1 0
Y vector in the associated pair no. 1 is:
1 1 0 0 1
Input vector is:

1 1 1 0 0 0
// We get here more of the detailed output as in the previous case. We will simply not present it
1 1 1 0 0 0
0 1 0 1 1
Input vector is:

0 1 1 0 0 0
A 0 is obtained, use previous output value
Program Output 172


0 1 1 0 0 0
1 0 0 1 0

0 1 1 0 0 0
1 0 0 1 0
Input vector is:

0 1 0 1 0 1

0 1 0 1 0 1
0 0 1 1 0

0 1 0 1 0 1
0 0 1 1 0
0 1 0 1 0 1
0 0 1 1 0
Input vector is:

1 1 1 1 1 1
Program Output 173


1 1 1 1 1 1
0 1 1 0 1
Program Output 174


1 0 0 1 1 1
0 1 1 0 1
1 0 0 1 1 1
0 1 1 0 1
THE FOLLOWING ASSOCIATED PAIRS WERE FOUND BY BAM
X vector in the associated pair no. 1 is: //first exemplar pair

1 0 1 0 1 0
1 1 0 0 1
X vector in the associated pair no. 2 is: //second exemplar pair

1 1 1 0 0 0
0 1 0 1 1
X vector in the associated pair no. 3 is: //third exemplar pair

0 1 1 0 0 0
1 0 0 1 0
X vector in the associated pair no. 4 is: //complement of X of the

0 1 0 1 0 1 first exemplar pair
Y vector in the associated pair no. 4 is: //complement of Y of the

0 0 1 1 0 first exemplar pair
X vector in the associated pair no. 5 is: //input was X = (1, 1, 1,

1 0 0 1 1 1 1, 1) but result was complement
of third exemplar pair
Y vector in the associated pair no. 5 is: with X of which Hamming
0 1 1 0 1 distance is the least.
Program Output 175


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Additional Issues
If you desire to use vectors with real number components, instead of binary numbers, you can do so. Your
model is then called a Continuous Bidirectional Associative Memory. A matrix W and its transpose are used
for the connection weights. However, the matrix W is not formulated as we described so far. The matrix is
arbitrarily chosen, and kept constant. The thresholding function is also chosen as a continuous function,
unlike the one used before. The changes in activations of neurons during training are according to extensions
the Cohen−Grossberg paradigm.
Michael A. Cohen and Stephen Grossberg developed their model for autoassociation, with a symmetric matrix
of weights. Stability is ensured by the Cohen−Grossberg theorem; there is no learning.
If, however, you desire the network to learn new pattern pairs, by modifying the weights so as to find
association between new pairs, you are designing what is called an Adaptive Bidirectional Associative
Memory (ABAM).
The law that governs the weight changes determines the type of ABAM you have, namely, the Competitive
ABAM, or the Differential Hebbian ABAM, or the Differential Competitive ABAM. Unlike in the ABAM
model, which is additive type, some products of outputs from the two layers, or the derivatives of the
threshold functions are used in the other models.
Here we present a brief description of a model, which is a variation of the BAM. It is called UBBAM
(Unipolar Binary Bidirectional Associative Memory).
Unipolar Binary Bidirectional Associative Memory

T. C. B. Yu and R. J. Mears describe a design of unipolar binary bidirectional associative memory, and its
implementation with a Smart Advanced Spatial Light Modulator (SASLM). The SASLM device is a
ferroelectric liquid crystal spatial light modulator. It is driven by a silicon CMOS backplane. We use their
paper to present some features of a unipolar binary bidirectional associative memory, and ignore the hardware
implementation of it.
Recall the procedure by which you determine the weight matrix W for a BAM network, as described in the
previous pages. As a first step, you convert each vector in the exemplar pairs into its bipolar version. If X and
Y are an exemplar pair (in bipolar versions), you take the product XT Y and add it to similar products from
other exemplar pairs, to get a weight matrix W. Some of the elements of the matrix W may be negative
numbers. In the unipolar context you do not have negative values and positive values at the same time. Only
one of them is allowed. Suppose you do not want any negative numbers; then one way of remedying the
situation is by adding a large enough constant to each element of the matrix. You cannot choose to add to only
the negative numbers that show up in the matrix. Let us look at an example.
Suppose you choose two pairs of vectors as possible exemplars. Let them be,
Additional Issues 176

X1 = (1, 0, 0, 1), Y1= (0, 1, 1)

and
X2 = (0, 1, 1, 0), Y2 = (1, 0, 1)
These you change into bipolar components and get, respectively, (1, –1, –1, 1), (–1, 1, 1), (–1, 1, 1, –1), and
(1, –1, 1). The calculation of W for the BAM was done as follows.
1 [−1 1 1] −1 [1 −1 1] −1 1 1 −1 1 −1 −2 2 0
W = −1 + 1 = 1 −1 −1+1 −1 1 = 2 −2 0
−1 1 1 −1 −1 1 −1 1 2 −2 0
1 −1 −1 1 1 −1 1 −1 −2 2 0
and
−2 2 2 −2
WT = 2 −2 −2 2
0 0 0 0
You see some negative numbers in the matrix W. If you add the constant m, m = 2, to all the elements in the
matrix, you get the following matrix, denoted by W~.
0 4 2
W~ = 4 0 2
4 0 2
0 4 2
You need a modification to the thresholding function as well. Earlier you had the following function.
1 if yj > 0 1 if xi > 0
bj|t+1 = bj|t if yj = 0 and ai|t+1 = ai|t if xi = 0
0 if yj < 0 0 if xi < 0
Now you need to change the right−hand sides of the conditions from 0 to the product of m and sum of the
inputs of the neurons, that is to m times the sum of the inputs. For brevity, let us use Si for the sum of the
inputs. This is not a constant, and its value depends on what values you get for neuron inputs in any one
direction. Then the thresholding function can be given as follows:
1 if yj > m Si 1 if xi > m Si
bj|t+1 = bj|t if yj = m Si and ai|t+1 = ai|t if xi = m Si
0 if yj < m Si 0 if xi < m Si
For example, the vector X1 = (1, 0, 0, 1) has the activity vector (0, 8, 4) and the corresponding output vector is
(0, 1, 1), which is, Y1 = (0, 1, 1). The value of Si is 2 = 1 + 0 + 0 + 1. Therefore, mSi = 4. The first component
of the activation vector is 0, which is smaller than 4, and so the first component of the output vector is 0. The
second component of the activation vector is 8, which is larger than 4, and so the second component of the
output vector is 1. The last component of the activity vector is 4, which is equal to mSi, and so you use the
third component of the vector Y1 = (0, 1, 1), namely 1, as the third component of the output.
Additional Issues 177


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
The example of Yu and Mears, consists of eight 8−component vectors in both directions of the Bidirectional
Associative Memory model. You can take these vectors, as they do, to define the pixel values in a 8x8 grid of
pixels. What is being accomplished is the association of two spatial patterns. In terms of the binary numbers
that show the pixel values, the patterns are shown in Figure 8.2. Call them Pattern I and Pattern II.
Figure 8.2 Two patterns, Pattern I and Pattern II, given by pixel values: 1 for black, 0 for white.
Instead of Pattern I, they used a corrupted form of it as given in Figure 8.3. There was no problem in the
network finding the associated pair (Pattern I, Pattern II).
Figure 8.3 Pattern I and corrupted Pattern I.
In Figure 8.3, the corrupted version of Pattern I differs from Pattern I, in 10 of the 64 places, a corruption of
15.6%. This corruption percentage is cited as the limit below which, Yu and Mears state, heteroassociative
recall is obtained. Thus noise elimination to a certain extent is possible with this model. As we have seen with
the Hopfield memory, an application of associative memory is pattern completion, where you are presented a
corrupted version of a pattern and you recall the true pattern. This is autoassociation. In the case of BAM, you
have heteroassociative re call with a corrupted input.
Summary
In this chapter, bidirectional associative memories are presented. The development of these memories is
largely due to Kosko. They share with Adaptive Resonance Theory the feature of resonance between the two
layers in the network. The bidirectional associative memories (BAM) network finds heteroassociation
between binary patterns, and these are converted to bipolar values to determine the connection weight matrix.
Even though there are connections in both directions between neurons in the two layers, essentially only one
weight matrix is involved. You use the transpose of this weight matrix for the connections in the opposite
direction. When one input at one end leads to some output at the other end, which in turn leads to output that
is the same as the previous input, resonance is reached and an associated pair is found.
The continuous bidirectional associative memory extends the binary model to the continuous case. Adaptive
bidirectional memories of different flavors are the result of incorporating different learning paradigms. A
unipolar binary version of BAM is also presented as an application of the BAM for pattern completion.
Summary 178
Summary 179

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Chapter 9
FAM: Fuzzy Associative Memory
Introduction
In this chapter, you we will learn about fuzzy sets and their elements, both for input and output of an
associative neural network. Every element of a fuzzy set has a degree of membership in the set. Unless this
degree of membership is 1, an element does not belong to the set (in the sense of elements of an ordinary set
belonging to the set). In a neural network of a fuzzy system the inputs, the outputs, and the connection weights
all belong fuzzily to the spaces that define them. The weight matrix will be a fuzzy matrix, and the activations
of the neurons in such a network have to be determined by rules of fuzzy logic and fuzzy set operations.
An expert system uses what are called crisp rules and applies them sequentially. The advantage in casting the
same problem in a fuzzy system is that the rules you work with do not have to be crisp, and the processing is
done in parallel. What the fuzzy systems can determine is a fuzzy association. These associations can be
modified, and the underlying phenomena better understood, as experience is gained. That is one of the reasons
for their growing popularity in applications. When we try to relate two things through a process of trial and
error, we will be implicitly and intuitively establishing an association that is gradually modified and perhaps
bettered in some sense. Several fuzzy variables may be present in such an exercise. That we did not have full
knowledge at the beginning is not a hindrance; there is some difference in using probabilities and using fuzzy
logic as well. The degree of membership assigned for an element of a set does not have to be as firm as the
assignment of a probability.
The degree of membership is, like a probability, a real number between 0 and 1. The closer it is to 1, the less
ambiguous is the membership of the element in the set concerned. Suppose you have a set that may or may
not contain three elements, say, a, b, and c. Then the fuzzy set representation of it would be by the ordered
triple (ma, mb, mc), which is called the fit vector, and its components are called fit values. For example, the
triple (0.5, 0.5, 0.5) shows that each of a, b, and c, have a membership equal to only one−half. This triple
itself will describe the fuzzy set. It can also be thought of as a point in the three−dimensional space. None of
such points will be outside the unit cube. When the number of elements is higher, the corresponding points
will be on or inside the unit hypercube.
It is interesting to note that this fuzzy set, given by the triple (0.5, 0.5, 0.5), is its own complement, something
that does not happen with regular sets. The complement is the set that shows the degrees of nonmembership.
The height of a fuzzy set is the maximum of its fit values, and the fuzzy set is said to be normal, if its height is
1. The fuzzy set with fit vector (0.3, 0.7, 0.4) has height 0.7, and it is not a normal fuzzy set. However, by
introducing an additional dummy component with fit value 1, we can extend it into a normal fuzzy set. The
desirability of normalcy of a fuzzy set will become apparent when we talk about recall in fuzzy associative
memories.
Chapter 9 FAM: Fuzzy Associative Memory 180

The subset relationship is also different for fuzzy sets from the way it is defined for regular sets. For example,
if you have a fuzzy set given by the triple (0.3, 0.7, 0.4), then any fuzzy set with a triple (a, b, c) such that a d
0.3, b d 0.7, and c d 0.4, is its fuzzy subset. For example, the fuzzy set given by the triple (0.1, 0.7, 0) is a
subset of the fuzzy set (0.3, 0.7, 0.4).
Association
Consider two fuzzy sets, one perhaps referring to the popularity of (or interest in) an exhibition and the other
to the price of admission. The popularity could be very high, high, fair, low, or very low. Accordingly, the fit
vector will have five components. The price of admission could be high, modest, or low. Its fit vector has
three components. A fuzzy associative memory system then has the association (popularity, price), and the
fuzzy set pair encodes this association.
We describe the encoding process and the recall in the following sections. Once encoding is completed,
associations between subsets of the first fuzzy set and subsets of the second fuzzy set are also established by
the same fuzzy associative memory system.
FAM Neural Network

The neural network for a fuzzy associative memory has an input and an output layer, with full connections in
both directions, just as in a BAM neural network. Figure 9.1 shows the layout. To avoid cluttering, the figure
shows the network with three neurons in field A and two neurons in field B, and only some of the connection
weights. The general case is analogous.
Figure 9.1 Fuzzy associative memory neural network.
This figure is the same as Figure 8.1 in Chapter 8, since the model is the same, except that
fuzzy vectors are used with the current model.
Encoding
Encoding for fuzzy associative memory systems is similar in concept to the encoding process for bidirectional
associative memories, with some differences in the operations involved. In the BAM encoding, bipolar
versions of binary vectors are used just for encoding. Matrices of the type XiT Yi are added together to get the
connection weight matrix. There are two basic operations with the elements of these vectors. They are
multiplication and addition of products. There is no conversion of the fit values before the encoding by the
fuzzy sets. The multiplication of elements is replaced by the operation of taking the minimum, and addition is
replaced by the operation of taking the maximum.
There are two methods for encoding. The method just described is what is called max–min composition. It is
used to get the connection weight matrix and also to get the outputs of neurons in the fuzzy associative
memory neural network. The second method is called correlation–product encoding. It is obtained the same
way as a BAM connection weight matrix is obtained. Max–min composition is what is frequently used in
practice, and we will confine our attention to this method.
Association 181
Association 182

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Example of Encoding
Suppose the two fuzzy sets we use to encode have the fit vectors ( 0.3, 0.7, 0.4, 0.2) and (0.4, 0.3, 0.9). Then
the matrix W is obtained by using max–min composition as follows.
0.3 [0.4 0.3 0.9] min(0.3,0.4) min(0.3,0.3) min(0.3,0.9) 0.3 0.3 0.3
W=0.7 =min(0.7,0.4) min(0.7,0.3) min(0.7,0.9)=0.4 0.3 0.7
0.4 min(0.4,0.4) min(0.4,0.3) min(0.4,0.9) 0.4 0.3 0.4
0.2 min(0.2,0.4) min(0.2,0.3) min(0.2,0.9) 0.2 0.2 0.2
Recall for the Example

If we input the fit vector (0.3, 0.7, 0.4, 0.2), the output (b1, b2, b3) is determined as follows, using bj = max(
min(a1, w1j), …, min(am, wmj), where m is the dimension of the ‘a’ fit vector, and wij is the ith row, jth column
element of the matrix W.
b1 = max(min(0.3, 0.3), min(0.7, 0.4), min(0.4, 0.4),

min(0.2, 0.2)) = max(0.3, 0.4, 0.4, 0.2) = 0.4
b2 = max(min(0.3, 0.3), min(0.7, 0.3), min(0.4, 0.3),
min(0.2, 0.2)) = max( 0.3, 0.3, 0.3, 0.2 ) = 0.3
b3 = max(min(0.3, 0.3), min(0.7, 0.7), min(0.4, 0.4),
min(0.2, 0.2)) = max (0.3, 0.7, 0.4, 0.2) = 0.7
The output vector (0.4, 0.3, 0.7) is not the same as the second fit vector used, namely (0.4, 0.3, 0.9), but it is a
subset of it, so the recall is not perfect. If you input the vector (0.4, 0.3, 0.7) in the opposite direction, using
the transpose of the matrix W, the output is (0.3, 0.7, 0.4, 0.2), showing resonance. If on the other hand you
input (0.4, 0.3, 0.9) at that end, the output vector is (0.3, 0.7, 0.4, 0.2), which in turn causes in the other
direction an output of (0.4, 0.3, 0.7) at which time there is resonance. Can we foresee these results? The
following section explains this further.
Recall
Let us use the operator o to denote max–min composition. Perfect recall occurs when the weight matrix is
obtained using the max–min composition of fit vectors U and V as follows:
(i) U o W = V if and only if height (U) e height (V).

(ii) V o WT = U if and only if height (V) e height (U).
Also note that if X and Y are arbitrary fit vectors with the same dimensions as U and V, then:
(iii) X o W 4 V.
(iv) Y o WT 4 U.
A 4 B is the notation to say A is a subset of B.
Example of Encoding 183

In the previous example, height of (0.3, 0.7, 0.4, 0.2) is 0.7, and height of (0.4, 0.3, 0.9) is 0.9. Therefore (0.4,
0.3, 0.9) as input, produced (0.3, 0.7, 0.4, 0.2) as output, but (0.3, 0.7, 0.4, 0.2) as input, produced only a
subset of (0.4, 0.3, 0.9). That both (0.4, 0.3, 0.7) and (0.4, 0.3, 0.9) gave the same output, (0.3, 0.7, 0.4, 0.2) is
in accordance with the corollary to the above, which states that if (X, Y) is a fuzzy associated memory, and if
X is a subset of X’, then (X’, Y) is also a fuzzy associated memory.
C++ Implementation
We use the classes we created for BAM implementation in C++, except that we call the neuron class
fzneuron, and we do not need some of the methods or functions in the network class. The header file, the
source file, and the output from an illustrative run of the program are given in the following. The header file is
called fuzzyam.hpp, and the source file is called fuzzyam.cpp.
Program details
The program details are analogous to the program details given in Chapter 8. The computations are done with
fuzzy logic. Unlike in the nonfuzzy version, a single exemplar fuzzy vector pair is used here. There are no
transformations to bipolar versions, since the vectors are fuzzy and not binary and crisp.
A neuron in the first layer is referred to as anrn, and the number of neurons in this layer is referred to as
anmbr. bnrn is the name we give to the array of neurons in the second layer and bnmbr denotes the size of
that array. The sequence of operations in the program are as follows:
• We ask the user to input the exemplar vector pair.

• We give the network the X vector, in the exemplar pair. We find the activations of the elements of
bnrn array and get corresponding output vector as a binary pattern. If this is the Y in the exemplar
pair, the network has made a desired association in one direction, and we go on to the next.step.
Otherwise we have a potential associated pair, one of which is X and the other is what we just got as
the output vector in the opposite layer. We say potential associated pair because we have the next step
to confirm the association.
• We run the bnrn array through the transpose of the weight matrix and calculate the outputs of the
anrn array elements. If, as a result, we get the vector X as the anrn array, we found an associated
pair, (X, Y). Otherwise, we repeat the two steps just described until we find an associated pair.
• We now work with the next pair of exemplar vectors in the same manner as above, to find an
associated pair.
• We assign serial numbers, denoted by the variable idn, to the associated pairs so we can print
them all together at the end of the program. The pair is called (X, Y) where X produces Y through the
weight matrix W, and Y produces X through the weight matrix, which is the transpose of W.
• A flag is used to have value 0 until confirmation of association is obtained, when the value of the
flag changes to 1.
• Functions compr1 and compr2 in the network class verify if the potential pair is indeed an
associated pair and set the proper value of the flag mentioned above.
• Functions comput1 and comput2 in the network class carry out the calculations to get the
activations and then find the output vector, in the respective directions of the fuzzy associative
memory network.
A lot of the code from the bidirectional associative memory (BAM) is used for the FAM. Here are the listings,
with comments added where there are differences between this code and the code for the BAM of Chapter 8.
C++ Implementation 184

C++ Implementation 185


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Header File
Listing 9.1 fuzzyam.h
//fuzzyam.h V. Rao, H. Rao
#define MXSIZ 10
class fzneuron
{
protected:
int nnbr;
int inn,outn;
float output;
float activation;
float outwt[MXSIZ];
char *name;
public:
fzneuron() { };
};
class exemplar
{
protected:
int xdim,ydim;
float v1[MXSIZ],v2[MXSIZ]; // this is different from BAM
public:
exemplar() { };
void getexmplr(int,int,float *,float *);
void prexmplr();
};
class asscpair
{
protected:
int xdim,ydim,idn;
float v1[MXSIZ],v2[MXSIZ];
public:
asscpair() { };
void getasscpair(int,int,int);
void prasscpair();
};
Header File 186

class potlpair
{
protected:
int xdim,ydim;
float v1[MXSIZ],v2[MXSIZ];
public:
potlpair() { };
void getpotlpair(int,int);
void prpotlpair();
};
class network
{
public:
int anmbr,bnmbr,flag,nexmplr,nasspr,ninpt;
fzneuron (anrn)[MXSIZ],(bnrn)[MXSIZ];
exemplar (e)[MXSIZ];
asscpair (as)[MXSIZ];
potlpair (pp)[MXSIZ];
float outs1[MXSIZ],outs2[MXSIZ]; // change from BAM to floats
double mtrx1[MXSIZ][MXSIZ],mtrx2[MXSIZ][MXSIZ]; // change from
BAM to doubles
network() { };
void getnwk(int,int,int,float [][6],float [][4]);
void prwts();
void iterate();
void findassc(float *);
void asgninpt(float *);
void asgnvect(int,float *,float *);
void comput1();
void comput2();
void prstatus();
};
Source File
Listing 9.2 fuzzyam.cpp
//fuzzyam.cpp V. Rao, H. Rao
#include "fuzzyam.h"
float max(float x,float y) //new for FAM

{
float u;
u = ((x>y) ? x : y );
return u;
}
float min(float x,float y) // new for FAM

{
float u;
u =( (x>y) ? y : x) ;
return u;
}
Source File 187

void fzneuron::getnrn(int m1,int m2,int m3,char *y)

{
int i;
name = y;
nnbr = m1;
outn = m2;
inn = m3;
outwt[i] = 0 ;
}
output = 0;
activation = 0;
}
void exemplar::getexmplr(int k,int l,float *b1,float *b2) // changed

from BAM
{
int i2;
xdim = k;
ydim = l;
for(i2=0;i2<xdim;++i2){
v1[i2] = b1[i2]; }
for(i2=0;i2<ydim;++i2){
v2[i2] = b2[i2]; }
void exemplar::prexmplr()
{
int i;
cout<<"\nX vector you gave is:\n";
cout<<v1[i]<<" ";}
cout<<"\nY vector you gave is:\n";
cout<<v2[i]<<" ";}
cout<<"\n";
}
void asscpair::getasscpair(int i,int j,int k)

{
idn = i;
xdim = j;
ydim = k;
}
void asscpair::prasscpair()
{
int i;
cout<<"\nX vector in the associated pair no. "<<idn<<"is:\n";
cout<<v1[i]<<" ";}
Source File 188

cout<<"\nY vector in the associated pair no. "<<idn<<"is:\n";
cout<<v2[i]<<" ";}
cout<<"\n";
}
void potlpair::getpotlpair(int k,int j)

{
xdim = k;
ydim = j;
void potlpair::prpotlpair()
{
int i;
cout<<"\nX vector in possible associated pair is:\n";
cout<<v1[i]<<" ";}
cout<<"\nY vector in possible associated pair is:\n";
cout<<v2[i]<<" ";}
cout<<"\n";
}
void network::getnwk(int k,int l,int k1,float b1[][6],float

b2[][4])
{
anmbr = k;
bnmbr = l;
nexmplr = k1;
nasspr = 0;
ninpt = 0;
int i,j,i2;
float tmp1,tmp2;
flag =0;
char *y1="ANEURON", *y2="BNEURON" ;
for(i=0;i<nexmplr;++i){
e[i].getexmplr(anmbr,bnmbr,b1[i],b2[i]);
e[i].prexmplr();
cout<<"\n";
}
anrn[i].fzneuron::getnrn(i,bnmbr,0,y1);}
bnrn[i].fzneuron::getnrn(i,0,anmbr,y2);}
tmp1 = 0.0;
Source File 189

tmp2 = min(e[i2].v1[i],e[i2].v2[j]);
tmp1 = max(tmp1,tmp2);
}
mtrx1[i][j] = tmp1;
mtrx2[j][i] = mtrx1[i][j];
}
prwts();
cout<<"\n";
}
void network::asgninpt(float *b)

{
int i,j;
cout<<"\n";
anrn[i].output = b[i];
outs1[i] = b[i];
}

{
int i;
break;
}
}

{
int i;
break;}
void network::comput1() //changed from BAM

{
int j;
int ii1;
float c1 =0.0,d1;
cout<<"\n";
Source File 190

d1 = min(outs1[ii1],mtrx1[ii1][j]);
c1 = max(c1,d1);
}
cout<<"\n output layer neuron "<<j<<" activation is "
<<c1<<"\n";
bnrn[j].output = bnrn[j].activation;
outs2[j] = bnrn[j].output;
cout<<"\n output layer neuron "<<j<<" output is "
<<bnrn[j].output<<"\n";
}
}
void network::comput2() //changed from BAM

{
int i;
int ii1;
float c1=0.0,d1;
for(ii1=0;ii1<bnmbr;++ii1){
d1 = min(outs2[ii1],mtrx2[ii1][i]);
c1 = max(c1,d1);}
anrn[i].activation = c1;
cout<<"\ninput layer neuron "<<i<<"activation is "
<<c1<<"\n";
anrn[i].output = anrn[i].activation;
outs1[i] = anrn[i].output;
cout<<"\n input layer neuron "<<i<<"output is "
<<anrn[i].output<<"\n";
}
void network::asgnvect(int j1,float *b1,float *b2)

{
int j2;
for(j2=0;j2<j1;++j2){
b2[j2] = b1[j2];}
void network::prwts()
{
int i3,i4;
cout<<"\n weights−− input layer to output layer: \n\n";
cout<<anrn[i3].outwt[i4]<<" ";}
cout<<"\n"; }
cout<<"\n";
cout<<"\nweights−− output layer to input layer: \n\n";
Source File 191

cout<<bnrn[i3].outwt[i4]<<" ";}
cout<<"\n"; }
cout<<"\n";
}
void network::iterate()
{
int i1;
findassc(e[i1].v1);
}
void network::findassc(float *b)

{
int j;
flag = 0;
asgninpt(b);
ninpt ++;
cout<<"\nInput vector is:\n" ;
for(j=0;j<6;++j){
cout<<b[j]<<" ";};
cout<<"\n";
pp[0].getpotlpair(anmbr,bnmbr);
asgnvect(anmbr,outs1,pp[0].v1);
comput1();
if(flag>=0){
asgnvect(bnmbr,outs2,pp[0].v2);
cout<<"\n";
pp[0].prpotlpair();
cout<<"\n";
comput2(); }
for(j=1;j<MXSIZ;++j){
pp[j].getpotlpair(anmbr,bnmbr);
asgnvect(anmbr,outs1,pp[j].v1);
comput1();
asgnvect(bnmbr,outs2,pp[j].v2);
pp[j].prpotlpair();
cout<<"\n";
compr1(j,j−1);
compr2(j,j−1);
if(flag == 0) {
Source File 192

int j2;
nasspr += 1;
j2 = nasspr;
as[j2].getasscpair(j2,anmbr,bnmbr);
asgnvect(anmbr,pp[j].v1,as[j2].v1);
asgnvect(bnmbr,pp[j].v2,as[j2].v2);
cout<<"\nPATTERNS ASSOCIATED:\n";
as[j2].prasscpair();
j = MXSIZ ;
}
else
if(flag == 1)
{
flag = 0;
comput1();
}
void network::prstatus()
{
int j;
cout<<"\nTHE FOLLOWING ASSOCIATED PAIRS WERE FOUND BY FUZZY AM\n\n";
for(j=1;j<=nasspr;++j){
as[j].prasscpair();
cout<<"\n";}
void main()
{
int ar = 6, br = 4, nex = 1;
float inptv[][6]={0.1,0.3,0.2,0.0,0.7,0.5,0.6,0.0,0.3,0.4,0.1,0.2};
float outv[][4]={0.4,0.2,0.1,0.0};
cout<<"\n\nTHIS PROGRAM IS FOR A FUZZY ASSOCIATIVE MEMORY NETWORK. THE
NETWORK \n";
cout<<"IS SET UP FOR ILLUSTRATION WITH "<<ar<<" INPUT NEURONS, AND "<<br;
cout<<" OUTPUT NEURONS.\n"<<nex<<" exemplars are used to encode \n";
static network famn;

famn.getnwk(ar,br,nex,inptv,outv);
famn.iterate();
famn.findassc(inptv[1]);
famn.prstatus();
Source File 193


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Output
The illustrative run of the previous program uses the fuzzy sets with fit vectors (0.1, 0.3, 0.2, 0.0, 0.7, 0.5) and
(0.4, 0.2, 0.1, 0.0). As you can expect according to the discussion earlier, recall is not perfect in the reverse
direction and the fuzzy associated memory consists of the pairs (0.1, 0.3, 0.2, 0.0, 0.4, 0.4) with (0.4, 0.2, 0.1,
0.0) and (0.1, 0.2, 0.2, 0, 0.2, 0.2) with (0.2, 0.2, 0.1, 0). The computer output is in such detail as to be
self−explanatory.
THIS PROGRAM IS FOR A FUZZY ASSOCIATIVE MEMORY NETWORK. THE NETWORK IS

SET UP FOR ILLUSTRATION WITH SIX INPUT NEURONS, AND FOUR OUTPUT NEURONS.
1 exemplars are used to encode

0.1 0.3 0.2 0 0.7 0.5
0.4 0.2 0.1 0
weights−−input layer to output layer:
0.1 0.1 0.1 0

0.3 0.2 0.1 0
0.2 0.2 0.1 0
0 0 0 0
0.4 0.2 0.1 0
0.4 0.2 0.1 0
weights−−output layer to input layer:
0.1 0.3 0.2 0 0.4 0.4

0.1 0.2 0.2 0 0.2 0.2
0.1 0.1 0.1 0 0.1 0.1
0 0 0 0 0 0
Input vector is:

0.1 0.3 0.2 0 0.7 0.5
output layer neuron 0 activation is 0.4
output layer neuron 0 output is 0.4
Output 194

0.1 0.3 0.2 0 0.7 0.5
0.4 0.2 0.1 0
input layer neuron 0 activation is 0.1
input layer neuron 0 output is 0.1

0.1 0.3 0.2 0 0.4 0.4
0.4 0.2 0.1 0

0.1 0.3 0.2 0 0.4 0.4
0.4 0.2 0.1 0
Input vector is:

0.6 0 0.3 0.4 0.1 0.2
Output 195

0.6 0 0.3 0.4 0.1 0.2
0.2 0.2 0.1 0
Output 196
0.1 0.2 0.2 0 0.2 0.2

0.2 0.2 0.1 0

0.1 0.2 0.2 0 0.2 0.2
0.2 0.2 0.1 0

0.1 0.2 0.2 0 0.2 0.2
0.2 0.2 0.1 0
THE FOLLOWING ASSOCIATED PAIRS WERE FOUND BY FUZZY AM

0.1 0.3 0.2 0 0.4 0.4
0.4 0.2 0.1 0

0.1 0.2 0.2 0 0.2 0.2
0.2 0.2 0.1 0
Output 197
Summary
In this chapter, bidirectional associative memories are presented for fuzzy subsets. The development of these
is largely due to Kosko. They share the feature of resonance between the two layers in the network with
Adaptive Resonance theory. Even though there are connections in both directions between neurons in the two
layers, only one weight matrix is involved. You use the transpose of this weight matrix for the connections in
the opposite direction. When one input at one end leads to some output at the other, which in turn leads to
output same as the previous input, resonance is reached and an associated pair is found. In the case of
bidirectional fuzzy associative memories, one pair of fuzzy sets determines one fuzzy associative memory
system. Fit vectors are used in max–min composition. Perfect recall in both directions is not the case unless
the heights of both fit vectors are equal. Fuzzy associative memories can improve the performance of an
expert system by allowing fuzzy rules.
Summary 198

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Chapter 10
Adaptive Resonance Theory (ART)
Introduction
Grossberg’s Adaptive Resonance Theory, developed further by Grossberg and Carpenter, is for the
categorization of patterns using the competitive learning paradigm. It introduces a gain control and a reset to
make certain that learned categories are retained even while new categories are learned and thereby addresses
the plasticity–stability dilemma.
Adaptive Resonance Theory makes much use of a competitive learning paradigm. A criterion is developed to
facilitate the occurrence of winner−take−all phenomenon. A single node with the largest value for the set
criterion is declared the winner within its layer, and it is said to classify a pattern class. If there is a tie for the
winning neuron in a layer, then an arbitrary rule, such as the first of them in a serial order, can be taken as the
winner.
The neural network developed for this theory establishes a system that is made up of two subsystems, one
being the attentional subsystem, and this contains the unit for gain control. The other is an orienting
subsystem, and this contains the unit for reset. During the operation of the network modeled for this theory,
patterns emerge in the attentional subsystem and are called traces of STM (short−term memory). Traces of
LTM (long−term memory) are in the connection weights between the input layer and output layer.
The network uses processing with feedback between its two layers, until resonance occurs. Resonance occurs
when the output in the first layer after feedback from the second layer matches the original pattern used as
input for the first layer in that processing cycle. A match of this type does not have to be perfect. What is
required is that the degree of match, measured suitably, exceeds a predetermined level, termed vigilance
parameter. Just as a photograph matches the likeness of the subject to a greater degree when the granularity is
higher, the pattern match gets finer when the vigilance parameter is closer to 1.
The Network for ART1

The neural network for the adaptive resonance theory or ART1 model consists of the following:
• A layer of neurons, called the F 1 layer (input layer or comparison layer)

• A node for each layer as a gain control unit
• A layer of neurons, called the F 2 layer (output layer or recognition layer)
• A node as a reset unit
• Bottom−up connections from F 1 layer to F2 layer
• Top−down connections from F 2 layer to F1 layer
• Inhibitory connection (negative weight) form F 2 layer to gain control
• Excitatory connection (positive weight) from gain control to a layer
Chapter 10 Adaptive Resonance Theory (ART) 199

• Inhibitory connection from F 1 layer to reset node

• Excitatory connection from reset node to F 2 layer
A Simplified Diagram of Network Layout
Figure 10.1 simplified diagram of the neural network for an ART1 model.
Processing in ART1
The ART1 paradigm, just like the Kohonen Self−Organizing Map to be introduced in Chapter 11, performs
data clustering on input data; like inputs are clustered together into a category. As an example, you can use a
data clustering algorithm such as ART1 for Optical Character Recognition (OCR), where you try to match
different samples of a letter to its ASCII equivalent. Particular attention is made in the ART1 paradigm to
ensure that old information is not thrown away while new information is assimilated.
An input vector, when applied to an ART1 system, is first compared to existing patterns in the system. If there
is a close enough match within a specified tolerance (as indicated by a vigilance parameter), then that stored
pattern is made to resemble the input pattern further and the classification operation is complete. If the input
pattern does not resemble any of the stored patterns in the system, then a new category is created with a new
stored pattern that resembles the input pattern.
Special Features of the ART1 Model
One special feature of an ART1 model is that a two−thirds rule is necessary to determine the activity of
neurons in the F1 layer. There are three input sources to each neuron in layer F1. They are the external input,
the output of gain control, and the outputs of F2 layer neurons. The F1neurons will not fire unless at least two
of the three inputs are active. The gain control unit and the two−thirds rule together ensure proper response
from the input layer neurons. A second feature is that a vigilance parameter is used to determine the activity of
the reset unit, which is activated whenever there is no match found among existing patterns during
classification.
Notation for ART1 Calculations
Let us list the various symbols we will use to describe the operation of a neural network for an ART1 model:
wij Weight on the connection from the ith neuron in the F1 layer to the jth neuron in the F2 layer
vji Weight on the connection from the jth neuron in the F2 layer to the ith neuron on the F1 layer
ai Activation of the ith neuron in the F1 layer
bj Activation of the jth neuron in the F2 layer
xi Output of the ith neuron in the F1 layer
yj Output of the jth neuron in the F2 layer
zi Input to the ith neuron in F1 layer from F2 layer
A Simplified Diagram of Network Layout 200

Á Vigilance parameter, positive and no greater than 1 (0<Á d 1)

m Number of neurons in the F1 layer
n Number of neurons in the F2 layer
I Input vector
Si Sum of the components of the input vector
Sx Sum of the outputs of neurons in the F1 layer
A, C, D Parameters with positive values or zero
L Parameter with value greater than 1
B Parameter with value less than D + 1 but at least as large as either D or 1
r Index of winner of competition in the F2 layer
A Simplified Diagram of Network Layout 201


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Algorithm for ART1 Calculations
The ART1 equations are not easy to follow. We follow the description of the algorithm found in James A.
Freeman and David M. Skapura. The following equations, taken in the order given, describe the steps in the
algorithm. Note that binary input patterns are used in ART1.
Initialization of Parameters
wij should be positive and less than L / ( m − 1 + L)

vji should be greater than ( B − 1 ) / D
ai = −B / ( 1 + C )
Equations for ART1 Computations
When you read below the equations for ART1 computations, keep in mind the following considerations. If a
subscript i appears on the left−hand side of the equation, it means that there are m such equations, as the
subscript i varies from 1 to m. Similarly, if instead a subscript j occurs, then there are n such equations as j
ranges from 1 to n. The equations are used in the order they are given. They give a step−by−step description
of the following algorithm. All the variables, you recall, are defined in the earlier section on notation. For
example, I is the input vector.
F1 layer calculations:
ai = Ii / ( 1 + A ( Ii + B ) + C )
xi = 1 if ai > 0
0 if ai d 0
bj = £ wij xi, the summation being on i from 1 to m

yj = 1 if jth neuron has the largest activation value in the F2
layer
= 0 if jth neuron is not the winner in F2 layer
Top−down inputs:
zi = £vjiyj, the summation being on j from 1 to n (You will

notice that exactly one term is nonzero)
Algorithm for ART1 Calculations 202

ai = ( Ii + D zi − B ) / ( 1 + A ( Ii + D zi ) + C )
xi = 1 if ai > 0
= 0 if ai d 0
Checking with vigilance parameter:
If ( Sx / SI ) <£, set yj = 0 for all j, including the winner r in F2 layer, and consider the jth neuron inactive (this
step is reset, skip remaining steps).
If ( Sx / SI ) e £, then continue.
Modifying top−down and bottom−up connection weight for winner r:
vir = ( L / ( Sx + L −1 ) if xi = 1
= 0 if xi = 0
wri = 1 if xi = 1
= 0 if xi = 0
Having finished with the current input pattern, we repeat these steps with a new input pattern. We lose the
index r given to one neuron as a winner and treat all neurons in the F2 layer with their original indices
(subscripts).
The above presentation of the algorithm is hoped to make all the steps as clear as possible. The process is
rather involved. To recapitulate, first an input vector is presented to the F1 layer neurons, their activations are
determined, and then the threshold function is used. The outputs of the F1 layer neurons constitute the inputs
to the F2 layer neurons, from which a winner is designated on the basis of the largest activation. The winner
only is allowed to be active, meaning that the output is 1 for the winner and 0 for all the rest. The equations
implicitly incorporate the use of the 2/3 rule that we mentioned earlier, and they also incorporate the way the
gain control is used. The gain control is designed to have a value 1 in the phase of determining the activations
of the neurons in the F2 layer and 0 if either there is no input vector or output from the F2 layer is propagated
to the F1 layer.
Other Models
Extensions of an ART1 model, which is for binary patterns, are ART2 and ART3. Of these, ART2 model
categorizes and stores analog−valued patterns, as well as binary patterns, while ART3 addresses
computational problems of hierarchies.
C++ Implementation
Again, the algorithm for ART1 processing as given in Freeman and Skapura is followed for our C++
implementation. Our objective in programming ART1 is to provide a feel for the workings of this paradigm
with a very simple program implementation. For more details on the inner workings of ART1, you are
encouraged to consult Freeman and Skapura, or other references listed at the back of the book.
A Header File for the C++ Program for the ART1 Model Network
The header file for the C++ program for the ART1 model network is art1net.hpp. It contains the declarations
for two classes, an artneuron class for neurons in the ART1 model, and a network class, which is declared as
a friend class in the artneuron class. Functions declared in the network class include one to do the iterations
for the network operation, finding the winner in a given iteration, and one to inquire if reset is needed.
Other Models 203

//art1net.h V. Rao, H. Rao

//Header file for ART1 model network program
#define MXSIZ 10
class artneuron
{
protected:
int nnbr;
int inn,outn;
int output;
double activation;
double outwt[MXSIZ];
char *name;
public:
artneuron() { };
};
class network
{
public:
int anmbr,bnmbr,flag,ninpt,sj,so,winr;
float ai,be,ci,di,el,rho;
artneuron (anrn)[MXSIZ],(bnrn)[MXSIZ];
int outs1[MXSIZ],outs2[MXSIZ];
int lrndptrn[MXSIZ][MXSIZ];
double acts1[MXSIZ],acts2[MXSIZ];
double mtrx1[MXSIZ][MXSIZ],mtrx2[MXSIZ][MXSIZ];
network() { };
void getnwk(int,int,float,float,float,float,float);
void prwts1();
void prwts2();
int winner(int k,double *v,int);
void practs1();
void practs2();
void prouts1();
void prouts2();
void iterate(int *,float,int);
void asgninpt(int *);
void comput1(int);
void comput2(int *);
void prlrndp();
void inqreset(int);
void adjwts1();
void adjwts2();
};
Other Models 204


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
A Source File for C++ Program for an ART1 Model Network
The implementations of the functions declared in the header file are contained in the source file for the C++
program for an ART1 model network. It also has the main function, which contains specifications of the
number of neurons in the two layers of the network, the values of the vigilance and other parameters, and the
input vectors. Note that if there are n neurons in a layer, they are numbered serially from 0 to n–1, and not
from 1 to n in the C++ program. The source file is called art1net.cpp. It is set up with six neurons in the F1
layer and seven neurons in the F2 layer. The main function also contains the parameters needed in the
algorithm.
To initialize the bottom−up weights, we set each weight to be –0.1 + L/(m – 1 + L) so that it is greater than 0
and less than L/(m – 1 + L), as suggested before. Similarly, the top−down weights are initialized by setting
each of them to 0.2 + (B – 1)/D so it would be greater than (B – 1)/D. Initial activations of the F1 layer
neurons are each set to –B/(1 + C), as suggested earlier.
A restrmax function is defined to compute the maximum in an array when one of the array elements is not
desired to be a candidate for the maximum. This facilitates the removal of the current winner from
competition when reset is needed. Reset is needed when the degree of match is of a smaller magnitude than
the vigilance parameter.
The function iterate is a member function of the network class and does the processing for the network. The
inqreset function of the network class compares the vigilance parameter with the degree of match.
//art1net.cpp V. Rao, H. Rao

//Source file for ART1 network program
#include "art1net.h"
int restrmax(int j,double *b,int k)

{
int i,tmp;
for(i=0;i<j;i++){
if(i !=k)
{tmp = i;
i = j;}
}
for(i=0;i<j;i++){
if( (i != tmp)&&(i != k))
{if(b[i]>b[tmp]) tmp = i;}}
return tmp;
}
A Source File for C++ Program for an ART1 Model Network 205
void artneuron::getnrn(int m1,int m2,int m3, char *y)

{
int i;
name = y;
nnbr = m1;
outn = m2;
inn = m3;
outwt[i] = 0 ;
}
output = 0;
activation = 0.0;
}
void network::getnwk(int k,int l,float aa,float bb,float

cc,float dd,float ll)
{
anmbr = k;
bnmbr = l;
ninpt = 0;
ai = aa;
be = bb;
ci = cc;
di = dd;
el = ll;
int i,j;
flag = 0;
char *y1="ANEURON", *y2="BNEURON" ;
anrn[i].artneuron::getnrn(i,bnmbr,0,y1);}
bnrn[i].artneuron::getnrn(i,0,anmbr,y2);}
float tmp1,tmp2,tmp3;
tmp1 = 0.2 +(be − 1.0)/di;
tmp2 = −0.1 + el/(anmbr − 1.0 +el);
tmp3 = − be/(1.0 + ci);
anrn[i].activation = tmp3;
acts1[i] = tmp3;
mtrx1[i][j] = tmp1;
mtrx2[j][i] = tmp2;
}
}
prwts1();
prwts2();
practs1();
cout<<"\n";
}
int network::winner(int k,double *v,int kk){

int t1;
t1 = restrmax(k,v,kk);
return t1;
}
void network::prwts1()
{
int i3,i4;
cout<<"\nweights for F1 layer neurons: \n";
cout<<anrn[i3].outwt[i4]<<" ";}
cout<<"\n"; }
cout<<"\n";
}
void network::prwts2()
{
int i3,i4;
cout<<"\nweights for F2 layer neurons: \n";
cout<<bnrn[i3].outwt[i4]<<" ";};
cout<<"\n"; }
cout<<"\n";
}
void network::practs1()
{
int j;
cout<<"\nactivations of F1 layer neurons: \n";
for(j=0;j<anmbr;++j){
cout<<acts1[j]<<" ";}
cout<<"\n";
}
void network::practs2()
{
int j;
cout<<"\nactivations of F2 layer neurons: \n";
cout<<acts2[j]<<" ";}
cout<<"\n";
}
void network::prouts1()
{
int j;
cout<<"\noutputs of F1 layer neurons: \n";
cout<<outs1[j]<<" ";}
cout<<"\n";
}
void network::prouts2()
{
int j;
cout<<"\noutputs of F2 layer neurons: \n";
cout<<outs2[j]<<" ";}
cout<<"\n";
}
void network::asgninpt(int *b)

{
int j;
sj = so = 0;
cout<<"\nInput vector is:\n" ;
cout<<b[j]<<" ";}
cout<<"\n";
sj += b[j];
anrn[j].activation = b[j]/(1.0 +ci +ai*(b[j]+be));
acts1[j] = anrn[j].activation;
if(anrn[j].activation > 0) anrn[j].output = 1;
else
anrn[j].output = 0;
outs1[j] = anrn[j].output;
so += anrn[j].output;
}
practs1();
prouts1();
}
void network::inqreset(int t1)

{
int jj;
flag = 0;
jj = so/sj;
cout<<"\ndegree of match: "<<jj<<" vigilance: "<<rho<<"\n";
if( jj > rho ) flag = 1;
else
{cout<<"winner is "<<t1;
cout<<" reset required \n";}
void network::comput1(int k)
{
int j;
int ii1;
double c1 = 0.0;
cout<<"\n";
c1 += outs1[ii1] * mtrx2[j][ii1];
}
acts2[j] = c1;};
winr = winner(bnmbr,acts2,k);
cout<<"winner is "<<winr;
if(j == winr) bnrn[j].output = 1;
else bnrn[j].output = 0;
outs2[j] = bnrn[j].output;
}
practs2();
prouts2();
}
void network::comput2(int *b)

{
double db[MXSIZ];
double tmp;
so = 0;
int i,j;
db[j] =0.0;
db[j] += mtrx1[j][i]*outs2[i];};
tmp = b[j] + di*db[j];

acts1[j] = (tmp − be)/(ci +1.0 +ai*tmp);
anrn[j].activation = acts1[j];
if(anrn[j].activation > 0) anrn[j].output = 1;
else anrn[j].output = 0;
outs1[j] = anrn[j].output;
so += anrn[j].output;
}
cout<<"\n";
practs1();
prouts1();
}
void network::adjwts1()
{
int i;
if(outs1[i] >0) {mtrx1[i][winr] = 1.0;}
else
{mtrx1[i][winr] = 0.0;}
anrn[i].outwt[winr] = mtrx1[i][winr];}
prwts1();
}
void network::adjwts2()
{
int i;
cout<<"\nwinner is "<<winr<<"\n";
if(outs1[i] > 0) {mtrx2[winr][i] = el/(so + el −1);}
else
{mtrx2[winr][i] = 0.0;}
bnrn[winr].outwt[i] = mtrx2[winr][i];}
prwts2();
}
void network::iterate(int *b,float rr,int kk)

{
int j;
rho = rr;
flag = 0;
asgninpt(b);
comput1(kk);
comput2(b);
inqreset(winr);
if(flag == 1){
ninpt ++;
adjwts1();
adjwts2();
int j3;
for(j3=0;j3<anmbr;++j3){
lrndptrn[ninpt][j3] = b[j3];}
prlrndp();
}
else
outs2[j] = 0;
bnrn[j].output = 0;}
iterate(b,rr,winr);
}
}
void network::prlrndp()
{
int j;
cout<<"\nlearned vector # "<<ninpt<<" :\n";
cout<<lrndptrn[ninpt][j]<<" ";}
cout<<"\n";
}
void main()
{
int ar = 6, br = 7, rs = 8;
float aa = 2.0,bb = 2.5,cc = 6.0,dd = 0.85,ll = 4.0,rr =
0.95;
int inptv[][6]={0,1,0,0,0,0,1,0,1,0,1,0,0,0,0,0,1,0,1,0,1,0,\
1,0};
cout<<"\n\nTHIS PROGRAM IS FOR AN −ADAPTIVE RESONANCE THEORY\

1 − NETWORK.\n";
cout<<"THE NETWORK IS SET UP FOR ILLUSTRATION WITH "<<ar<<" \
INPUT NEURONS,\n";
cout<<" AND "<<br<<" OUTPUT NEURONS.\n";
static network bpn;

bpn.getnwk(ar,br,aa,bb,cc,dd,ll) ;
bpn.iterate(inptv[0],rr,rs);
}

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Program Output
Four input vectors are used in the trial run of the program, and these are specified in the main function. The
output is self−explanatory. We have included only in this text some comments regarding the output. These
comments are enclosed within strings of asterisks. They are not actually part of the program output. Table
10.1 shows a summarization of the categorization of the inputs done by the network. Keep in mind that the
numbering of the neurons in any layer, which has n neurons, is from 0 to n – 1, and not from 1 to n.
Table 10.1 Categorization of Inputs

inputwinner in F2 layer
0 1 0 0 0 00, no reset
1 0 1 0 1 01, no reset
0 0 0 0 1 01, after reset 2
1 0 1 0 1 01, after reset 3
The input pattern 0 0 0 0 1 0 is considered a subset of the pattern 1 0 1 0 1 0 in the sense that in whatever
position the first pattern has a 1, the second pattern also has a 1. Of course, the second pattern has 1’s in other
positions as well. At the same time, the pattern 1 0 1 0 1 0 is considered a superset of the pattern 0 0 0 0 1 0.
The reason that the pattern 1 0 1 0 1 0 is repeated as input after the pattern 0 0 0 0 1 0 is processed, is to see
what happens with this superset. In both cases, the degree of match falls short of the vigilance parameter, and
a reset is needed.
Here’s the output of the program:
THIS PROGRAM IS FOR AN ADAPTIVE RESONANCE THEORY

1−NETWORK. THE NETWORK IS SET UP FOR ILLUSTRATION WITH SIX INPUT NEURONS
AND SEVEN OUTPUT NEURONS.
*************************************************************
Initialization of connection weights and F1 layer activations. F1 layer
connection weights are all chosen to be equal to a random value subject
to the conditions given in the algorithm. Similarly, F2 layer connection
weights are all chosen to be equal to a random value subject to the
conditions given in the algorithm.
*************************************************************
weights for F1 layer neurons:
1.964706 1.964706 1.964706 1.964706 1.964706 1.964706 1.964706
1.964706 1.964706 1.964706 1.964706 1.964706 1.964706 1.964706
1.964706 1.964706 1.964706 1.964706 1.964706 1.964706 1.964706
1.964706 1.964706 1.964706 1.964706 1.964706 1.964706 1.964706
1.964706 1.964706 1.964706 1.964706 1.964706 1.964706 1.964706
1.964706 1.964706 1.964706 1.964706 1.964706 1.964706 1.964706

0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
Program Output 213

0.344444 0.344444 0.344444 0.344444 0.344444 0.344444

0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
activations of F1 layer neurons:

−0.357143 −0.357143 −0.357143 −0.357143 −0.357143 −0.357143
*************************************************************
A new input vector and a new iteration
*************************************************************
Input vector is:
0 1 0 0 0 0

0 0.071429 0 0 0 0
outputs of F1 layer neurons:

0 1 0 0 0 0
winner is 0
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444 0.344444

1 0 0 0 0 0 0

−0.080271 0.013776 −0.080271 −0.080271 −0.080271 −0.080271

0 1 0 0 0 0
*************************************************************
Top−down and bottom−up outputs at F1 layer match, showing resonance.
*************************************************************
degree of match: 1 vigilance: 0.95

0 1.964706 1.964706 1.964706 1.964706 1.964706 1.964706
1 1.964706 1.964706 1.964706 1.964706 1.964706 1.964706
0 1.964706 1.964706 1.964706 1.964706 1.964706 1.964706
0 1.964706 1.964706 1.964706 1.964706 1.964706 1.964706
0 1.964706 1.964706 1.964706 1.964706 1.964706 1.964706
0 1.964706 1.964706 1.964706 1.964706 1.964706 1.964706
winner is 0

0 1 0 0 0 0
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
learned vector # 1 :
0 1 0 0 0 0
*************************************************************
*************************************************************
Input vector is:
1 0 1 0 1 0
Program Output 214


0.071429 0 0.071429 0 0.071429 0

1 0 1 0 1 0
winner is 1
0 1.033333 1.033333 1.033333 1.033333 1.033333 1.033333

0 1 0 0 0 0 0

0.013776 −0.080271 0.013776 −0.080271 0.013776 −0.080271

1 0 1 0 1 0
*************************************************************
Top−down and bottom−up outputs at F1 layer match,
showing resonance.
*************************************************************

0 1 1.964706 1.964706 1.964706 1.964706 1.964706
1 0 1.964706 1.964706 1.964706 1.964706 1.964706
0 1 1.964706 1.964706 1.964706 1.964706 1.964706
0 0 1.964706 1.964706 1.964706 1.964706 1.964706
0 1 1.964706 1.964706 1.964706 1.964706 1.964706
0 0 1.964706 1.964706 1.964706 1.964706 1.964706
winner is 1

0 1 0 0 0 0
0.666667 0 0.666667 0 0.666667 0
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
1 0 1 0 1 0
*************************************************************
*************************************************************
Input vector is:
0 0 0 0 1 0

0 0 0 0 0.071429 0

0 0 0 0 1 0
winner is 1
0 0.666667 0.344444 0.344444 0.344444 0.344444 0.344444
Program Output 215


0 1 0 0 0 0 0

−0.189655 −0.357143 −0.189655 −0.357143 −0.060748 −0.357143

0 0 0 0 0 0

winner is 1 reset required
*************************************************************
Input vector repeated after reset, and a new iteration
*************************************************************
Input vector is:
0 0 0 0 1 0

0 0 0 0 0.071429 0

0 0 0 0 1 0
winner is 2
0 0.666667 0.344444 0.344444 0.344444 0.344444 0.344444
0 0 1 0 0 0 0

−0.080271 −0.080271 −0.080271 −0.080271 0.013776 −0.080271

0 0 0 0 1 0
*************************************************************
Top−down and bottom−up outputs at F1 layer match, showing resonance.
*************************************************************

0 1 0 1.964706 1.964706 1.964706 1.964706
1 0 0 1.964706 1.964706 1.964706 1.964706
0 1 0 1.964706 1.964706 1.964706 1.964706
0 0 0 1.964706 1.964706 1.964706 1.964706
0 1 1 1.964706 1.964706 1.964706 1.964706
0 0 0 1.964706 1.964706 1.964706 1.964706
winner is 2

0 1 0 0 0 0
0.666667 0 0.666667 0 0.666667 0
0 0 0 0 1 0
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0 0 0 0 1 0
*************************************************************
An old (actually the second above) input vector is retried after trying a
Program Output 216

subset vector, and a new iteration

*************************************************************
Input vector is:
1 0 1 0 1 0

0.071429 0 0.071429 0 0.071429 0

1 0 1 0 1 0
winner is 1
0 2 1 1.033333 1.033333 1.033333 1.03333

0 1 0 0 0 0 0

−0.060748 −0.357143 −0.060748 −0.357143 −0.060748 −0.357143

0 0 0 0 0 0

winner is 1 reset required
*************************************************************
Input vector repeated after reset, and a new iteration
*************************************************************
Input vector is:
1 0 1 0 1 0

0.071429 0 0.071429 0 0.071429 0

1 0 1 0 1 0
winner is 3
0 2 1 1.033333 1.033333 1.033333 1.033333

0 0 0 1 0 0 0

0.013776 −0.080271 0.013776 −0.080271 0.013776 −0.080271

1 0 1 0 1 0
*************************************************************
Top−down and Bottom−up outputs at F1layer match, showing resonance.
*************************************************************

0 1 0 1 1.964706 1.964706 1.964706
1 0 0 0 1.964706 1.964706 1.964706
0 1 0 1 1.964706 1.964706 1.964706
0 0 0 0 1.964706 1.964706 1.964706
0 1 1 1 1.964706 1.964706 1.964706
0 0 0 0 1.964706 1.964706 1.964706
Program Output 217

winner is 3

0 1 0 0 0 0
0.666667 0 0.666667 0 0.666667 0
0 0 0 0 1 0
0.666667 0 0.666667 0 0.666667 0
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
0.344444 0.344444 0.344444 0.344444 0.344444 0.344444
1 0 1 0 1 0
Program Output 218


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Summary
This chapter presented the basics of the Adaptive Resonance Theory of Grossberg and Carpenter and a C++
implementation of the neural network modeled for this theory. It is an elegant theory that addresses the
stability–plasticity dilemma. The network relies on resonance. It is a self−organizing network and does
categorization by associating individual neurons of the F2 layer with individual patterns. By employing a
so−called 2/3 rule, it ensures stability in learning patterns.
Summary 219

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Chapter 11
The Kohonen Self−Organizing Map
Introduction
This chapter discusses one type of unsupervised competitive learning, the Kohonen feature map, or
self−organizing map (SOM). As you recall, in unsupervised learning there are no expected outputs presented
to a neural network, as in a supervised training algorithm such as backpropagation. Instead, a network, by its
self−organizing properties, is able to infer relationships and learn more as more inputs are presented to it. One
advantage to this scheme is that you can expect the system to change with changing conditions and inputs.
The system constantly learns. The Kohonen SOM is a neural network system developed by Teuvo Kohonen
of Helsinki University of Technology and is often used to classify inputs into different categories.
Applications for feature maps can be traced to many areas, including speech recognition and robot motor
control.
Competitive Learning
A Kohonen feature map may be used by itself or as a layer of another neural network. A Kohonen layer is
composed of neurons that compete with each other. Like in Adaptive Resonance Theory, the Kohonen SOM
is another case of using a winner−take−all strategy. Inputs are fed into each of the neurons in the Kohonen
layer (from the input layer). Each neuron determines its output according to a weighted sum formula:
Output = £ wij xi
The weights and the inputs are usually normalized, which means that the magnitude of the weight and input
vectors are set equal to one. The neuron with the largest output is the winner. This neuron has a final output of
1. All other neurons in the layer have an output of zero. Differing input patterns end up firing different winner
neurons. Similar or identical input patterns classify to the same output neuron. You get like inputs clustered
together. In Chapter 12, you will see the use of a Kohonen network in pattern classification.
Normalization of a Vector
Consider a vector, A = ax + by + cz. The normalized vector A’ is obtained by dividing each component of A
by the square root of the sum of squares of all the components. In other words each component is multiplied
by 1/ [radic](a2 + b2 + c2). Both the weight vector and the input vector are normalized during the operation of
the Kohonen feature map. The reason for this is the training law uses subtraction of the weight vector from the
input vector. Using normalization of the values in the subtraction reduces both vectors to a unit−less status,
and hence, makes the subtraction of like quantities possible. You will learn more about the training law
shortly.
Chapter 11 The Kohonen Self−Organizing Map 220

Lateral Inhibition
Lateral inhibition is a process that takes place in some biological neural networks. Lateral connections of
neurons in a given layer are formed, and squash distant neighbors. The strength of connections is inversely
related to distance. The positive, supportive connections are termed as excitatory while the negative,
squashing connections are termed inhibitory.
A biological example of lateral inhibition occurs in the human vision system.
The Mexican Hat Function
Figure 11.1 shows a function, called the mexican hat function, which shows the relationship between the
connection strength and the distance from the winning neuron. The effect of this function is to set up a
competitive environment for learning. Only winning neurons and their neighbors participate in learning for a
given input pattern.
Figure 11.1 The mexican hat function showing lateral inhibition.
Training Law for the Kohonen Map

The training law for the Kohonen feature map is straightforward. The change in weight vector for a given
output neuron is a gain constant, alpha, multiplied by the difference between the input vector and the old
weight vector:
Wnew = Wold + alpha * (Input −Wold)
Both the old weight vector and the input vector are normalized to unit length. Alpha is a gain constant
between 0 and 1.
Significance of the Training Law
Let us consider the case of a two−dimensional input vector. If you look at a unit circle, as shown in Figure
11.2, the effect of the training law is to try to align the weight vector and the input vector. Each pattern
attempts to nudge the weight vector closer by a fraction determined by alpha. For three dimensions the surface
becomes a unit sphere instead of a circle. For higher dimensions you term the surface a hypersphere. It is not
necessarily ideal to have perfect alignment of the input and weight vectors. You use neural networks for their
ability to recognize patterns, but also to generalize input data sets. By aligning all input vectors to the
corresponding winner weight vectors, you are essentially memorizing the input data set classes. It may be
more desirable to come close, so that noisy or incomplete inputs may still trigger the correct classification.
Figure 11.2 The training law for the Kohonen map as shown on a unit circle.
The Mexican Hat Function 221

The Neighborhood Size and Alpha
In the Kohonen map, a parameter called the neighborhood size is used to model the effect of the mexican hat
function. Those neurons that are within the distance specified by the neighborhood size participate in training
and weight vector changes; those that are outside this distance do not participate in learning. The
neighborhood size typically is started as an initial value and is decreased as the input pattern cycles continue.
This process tends to support the winner−take−all strategy by eventually singling out a winner neuron for a
given pattern.
Figure 11.3 shows a linear arrangement of neurons with a neighborhood size of 2. The hashed central neuron
is the winner. The darkened adjacent neurons are those that will participate in training.
Figure 11.3 Winner neuron with a neighborhood size of 2 for a Kohonen map.
Besides the neighborhood size, alpha typically is also reduced during simulation. You will see these features
when we develop a Kohonen map program.
The Neighborhood Size and Alpha 222


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
C++ Code for Implementing a Kohonen Map

The C++ code for the Kohonen map draws on much of the code developed for the backpropagation simulator.
The Kohonen map is a much simpler program and may not rely on as large a data set for input. The Kohonen
map program uses only two files, an input file and an output file. In order to use the program, you must create
an input data set and save this in a file called input.dat. The output file is called kohonen.dat and is saved in
your current working directory. You will get more details shortly on the formats of these files.
The Kohonen Network

The Kohonen network has two layers, an input layer and a Kohonen output layer. (See Figure 11.4). The input
layer is a size determined by the user and must match the size of each row (pattern) in the input data file.
Figure 11.4 A Kohonen network.
Modeling Lateral Inhibition and Excitation
The mexican hat function shows positive values for an immediate neighborhood around the neuron and
negative values for distant neurons. A true method of modeling would incorporate mutual excitation or
support for neurons that are within the neighborhood (with this excitation increasing for nearer neurons) and
inhibition for distant neurons outside the neighborhood. For the sake of computational efficiency, we model
lateral inhibition and excitation by looking at the maximum output for the output neurons and making that
output belong to a winner neuron. Other outputs are inhibited by setting their outputs to zero. Training, or
weight update, is performed on all outputs that are within a neighborhood size distance from the winner
neuron. Neurons outside the neighborhood do not participate in training. The true way of modeling lateral
inhibition would be too expensive since the number of lateral connections is quite large. You will find that
this approximation will lead to a network with many if not all of the properties of a true modeling approach of
a Kohonen network.
Classes to be Used
We use many of the classes from the backpropagation simulator. We require only two layers, the input layer
and the Kohonen layer. We make a new layer class called the Kohonen layer class, and a new network class
called the Kohonen_network.
C++ Code for Implementing a Kohonen Map 223

Revisiting the Layer Class
The layer class needs to be slightly modified, as shown in Listing 11.1.
Listing 11.1 Modification of layer.h

class network;
class Kohonen_network;
class layer
{
protected:
int num_inputs;
int num_outputs;
float *outputs;// pointer to array of outputs
friend network;
friend Kohonen_network; // update for Kohonen model
public:

};
...
Here the changes are indicated in italic. You notice that the Kohonen_network is made a friend to the layer
class, so that the Kohonen_network can have access to the data of a layer.
A New Layer Class for a Kohonen Layer
The next step to take is to create a Kohonen_layer class and a Kohonen_network class. This is shown in
Listing 11.2.
Listing 11.2 The Kohonen_layer class and Kohonen_network class in layerk.h
// layerk.h V.Rao, H. Rao

// header file for the Kohonen layer and
// the Kohonen network
class Kohonen_layer: public layer

{
protected:
float * weights;
Revisiting the Layer Class 224

int winner_index;
float win_distance;
int neighborhood_size;
friend Kohonen_network;
public:
Kohonen_layer(int, int, int);

~Kohonen_layer();
void update_neigh_size(int);
float get_win_dist();
};
class Kohonen_network
private:
layer *layer_ptr[2];
int layer_size[2];
public:
Kohonen_network();
~Kohonen_network();
void set_up_network(int);
void get_next_vector(FILE *);
void process_next_pattern();
int get_win_index();
};
The Kohonen_layer is derived from the layer class, so it has pointers inherited that point to a set of outputs
and a set of inputs. Let’s look at some of the functions and member variables.


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Kohonen_layer:
• float * weights Pointer to the weights matrix.

• int winner_index Index value of the output, which is the winner.
• float win_distance The Euclidean distance of the winner weight vector from the input vector.
• int neighborhood_size The size of the neighborhood.
• Kohonen_layer(int, int, int) Constructor for the layer: inputs, outputs, and the neighborhood
size.
• ~Kohonen_layer() Destructor.
• virtual void calc_out() The function to calculate the outputs; for the Kohonen layer this models
lateral competition.
• void randomize_weights() A function to initialize weights with random normal values.
• void update_neigh_size(nt) This function updates the neighborhood size with a new value.
• void update_weights(const float) This function updates the weights according to the training law
using the passed parameter, alpha.
• void list_weights() This function can be used to list the weight matrix.
• void list_outputs() This function is used to write outputs to the output file.
• float get_win_dist() Returns the Euclidean distance between the winner weight vector and the
input vector.
Kohonen_network:
• layer *layer_ptr[2] Pointer array; element 0 points to the input layer, element 1 points to the
Kohonen layer.
• int layer_size[2] Array of layer sizes for the two layers.
• int neighborhood_size The current neighborhood size.
• Kohonen_network() Constructor.
• ~Kohonen_network() Destructor.
• void get_layer_info() Gets information about the layer sizes.
• void set_up_network(int) Connects layers and sets up the Kohonen map.
• void randomize_weights() Creates random normalized weights.
• void update_neigh_size(int) Changes the neighborhood size.
• void update_weights(const float) Performs weight update according to the training law.
• void list_weights() Can be used to list the weight matrix.
• void list_outputs() Can be used to list outputs.
• void get_next_vector(FILE *) Function gets another input vector from the input file.
• void process_next_pattern() Applies pattern to the Kohonen map.
• float get_win_dist() Returns the winner’s distance from the input vector.
• int get_win_index() Returns the index of the winner.

Implementation of the Kohonen Layer and Kohonen Network

Listing 11.3 shows the layerk.cpp file, which has the implementation of the functions outlined.
Listing 11.3 Implementation file for the Kohonen layer and Kohonen network :layerk.cpp
// layerk.cpp V.Rao, H.Rao

#include “layer.cpp”
#include “layerk.h”
// ————————————————————−
// Kohonen layer
//—————————————————————
Kohonen_layer::Kohonen_layer(int i, int o, int

init_neigh_size)
{
num_inputs=i;
num_outputs=o;
neighborhood_size=init_neigh_size;
}
Kohonen_layer::~Kohonen_layer()
{
}
void Kohonen_layer::calc_out()
{
// implement lateral competition
// choose the output with the largest
// value as the winner; neighboring
// outputs participate in next weight
// update. Winner’s output is 1 while
// all other outputs are zero
int i,j,k;
float maxval;
winner_index=0;
maxval=−1000000;

{
{
k=i*num_outputs;
{
cout << “weights are blowing up\n”;
cout << “try a smaller learning constant\n”;
cout << “e.g. beta=0.02
aborting...\n”;
exit(1);
}
Implementation of the Kohonen Layer and Kohonen Network 227

}
// no squash function
outputs[j]=accumulator;
if (outputs[j] > maxval)
{
maxval=outputs[j];
winner_index=j;
}
accumulator=0;
}
// set winner output to 1

outputs[winner_index]=1.0;
// now zero out all other outputs
for (j=0; j< winner_index; j++)
outputs[j]=0;
for (j=num_outputs−1; j>winner_index; j—)
outputs[j]=0;
void Kohonen_layer::randomize_weights()
{
int i, j, k;

float discard;
float norm;

{
k=i*num_outputs;
{
}
}
// now need to normalize the weight vectors

// to unit length
// a weight vector is the set of weights for
// a given output

{
norm=0;
{
k=i*num_outputs;
norm+=weights[k+j]*weights[k+j];
}
norm = 1/((float)sqrt((double)norm));

{
k=i*num_outputs;
weights[k+j]*=norm;

}
}
void Kohonen_layer::update_neigh_size(int new_neigh_size)

{
neighborhood_size=new_neigh_size;
}
void Kohonen_layer::update_weights(const float alpha)

{
int i, j, k;
int start_index, stop_index;
// alpha*(input−weight)
// zero change if input and weight
// vectors are aligned
// only update those outputs that
// are within a neighborhood’s distance
// from the last winner
start_index = winner_index −
neighborhood_size;
if (start_index < 0)
start_index =0;
stop_index = winner_index +
neighborhood_size;
if (stop_index > num_outputs−1)

stop_index = num_outputs−1;

{
k=i*num_outputs;
for (j=start_index; j<=stop_index; j++)
weights[k+j] +=
alpha*((*(inputs+i))−weights[k+j]);
}
void Kohonen_layer::list_weights()
{
int i, j, k;

{
k=i*num_outputs;
cout << “weight[“<<i<<”,”<<
j<<”] is: “<<weights[k+j];
}
}
void Kohonen_layer::list_outputs()
{
int i;
for (i=0; i< num_outputs; i++)

{

cout << “outputs[“<<i<<

“] is: “<<outputs[i];
}
float Kohonen_layer::get_win_dist()
{
int i, j, k;
j=winner_index;
float * win_dist_vec = new float [num_inputs];

{
k=i*num_outputs;
win_dist_vec[i]=(*(inputs+i))−weights[k+j];
accumulator+=win_dist_vec[i]*win_dist_vec[i];
}
win_distance =(float)sqrt((double)accumulator);
delete [num_inputs]win_dist_vec;
return win_distance;
Kohonen_network::Kohonen_network()
{
Kohonen_network::~Kohonen_network()
{
}
void Kohonen_network::get_layer_info()
{
int i;
//—————————————————————
//
// Get layer sizes for the Kohonen network
//
// ————————————————————−
cout << “ Enter in the layer sizes separated by spaces.\n”;

cout << “ A Kohonen network has an input layer \n”;
cout << “ followed by a Kohonen (output) layer \n”;
for (i=0; i<2; i++)

{
}
// ———————————————————————————
// size of layers:
// Kohonen_layer layer_size[1]
//———————————————————————————−

}
void Kohonen_network::set_up_network(int nsz)
{
int i;
// set up neighborhood size

neighborhood_size = nsz;
//———————————————————————————−
//
//———————————————————————————−
layer_ptr[1] =
new Kohonen_layer(layer_size[0],
layer_size[1],neighborhood_size);
for (i=0;i<2;i++)
{
{
cout << “insufficient memory\n”;
cout << “use a smaller architecture\n”;
exit(1);
}
}
//———————————————————————————−
//
//———————————————————————————−
// set inputs to previous layer outputs for the Kohonen layer
layer_ptr[1]−>inputs = layer_ptr[0]−>outputs;
void Kohonen_network::randomize_weights()
{
((Kohonen_layer *)layer_ptr[1])
}
void Kohonen_network::update_neigh_size(int n)
{
−>update_neigh_size(n);
}
void Kohonen_network::update_weights(const float a)

{
−>update_weights(a);
}
void Kohonen_network::list_weights()
{
−>list_weights();

void Kohonen_network::list_outputs()
{
−>list_outputs();
}
void Kohonen_network::get_next_vector(FILE * ifile)

{
int i;
float normlength=0;
int num_inputs=layer_ptr[1]−>num_inputs;
float *in = layer_ptr[1]−>inputs;
// get a vector and normalize it
{
fscanf(ifile,”%f”,(in+i));
normlength += (*(in+i))*(*(in+i));
}
fscanf(ifile,”\n”);
normlength = 1/(float)sqrt((double)normlength);
{
(*(in+i)) *= normlength;
}
void Kohonen_network::process_next_pattern()
{
layer_ptr[1]−>calc_out();
}
float Kohonen_network::get_win_dist()
{
float retval;
retval=((Kohonen_layer *)layer_ptr[1])
−>get_win_dist();
return retval;
}
int Kohonen_network::get_win_index()
{
return ((Kohonen_layer *)layer_ptr[1])
−>winner_index;
}


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Flow of the Program and the main() Function

The main() function is contained in a file called kohonen.cpp, which is shown in Listing 11.4. To compile
this program, you need only compile and make this main file, kohonen.cpp. Other files are included in this.
Listing 11.4 The main implementation file, kohonen.cpp for the Kohonen Map program
// kohonen.cpp V. Rao, H. Rao

// Program to simulate a Kohonen map
#include “layerk.cpp”
#define INPUT_FILE “input.dat”

#define OUTPUT_FILE “kohonen.dat”
#define dist_tol 0.05
void main()
{
int neighborhood_size, period;

float avg_dist_per_cycle=0.0;
float dist_last_cycle=0.0;
float avg_dist_per_pattern=100.0; // for the latest cycle
float dist_last_pattern=0.0;
float total_dist;
float alpha;
unsigned startup;
int max_cycles;
int patterns_per_cycle=0;
int total_cycles, total_patterns;

Kohonen_network knet;
FILE * input_file_ptr, * output_file_ptr;
// open input file for reading

if ((input_file_ptr=fopen(INPUT_FILE,”r”))==NULL)
{
cout << “problem opening input file\n”;
exit(1);
}
// open writing file for writing

if ((output_file_ptr=fopen(OUTPUT_FILE,”w”))==NULL)
{
cout << “problem opening output file\n”;
exit(1);
}
Flow of the Program and the main() Function 233

// ————————————————————−
// Read in an initial values for alpha, and the
// neighborhood size.
// Both of these parameters are decreased with
// time. The number of cycles to execute before
// decreasing the value of these parameters is
// called the period. Read in a value for the
// period.
// ————————————————————−
cout << “ Please enter initial values for:\n”;
cout << “alpha (0.01−1.0),\n”;
cout << “and the neighborhood size (integer between 0
and
50)\n”;
cout << “separated by spaces, e.g. 0.3 5 \n “;
cin >> alpha >> neighborhood_size ;
cout << “\nNow enter the period, which is the\n”;

cout << “number of cycles after which the values\n”;
cout << “for alpha the neighborhood size are
decremented\n”;
cout << “choose an integer between 1 and 500 , e.g. 50 \n”;
cin >> period;

cout << “\nPlease enter the maximum cycles for the
simulation\n”;
cout << “A cycle is one pass through the data set.\n”;
cout << “Try a value of 500 to start with\n\n”;
cin >> max_cycles;
// the main loop

//
// continue looping until the average distance is less
// than the tolerance specified at the top of this file
// , or the maximum number of
// cycles is exceeded;

knet.get_layer_info();

knet.set_up_network(neighborhood_size);
// randomize weights for the Kohonen layer

// note that the randomize function for the
// Kohonen simulator generates
// weights that are normalized to length = 1
knet.randomize_weights();
// write header to output file

fprintf(output_file_ptr,
“cycle\tpattern\twin index\tneigh_size\tavg_dist_per_pa

tern\n”);
“———————————————————————————\n”);
// main loop
startup=1;
total_dist=0;
while (
(avg_dist_per_pattern > dist_tol)
|| (startup==1)
)
{
startup=0;
dist_last_cycle=0; // reset for each cycle
while (!feof(input_file_ptr))
{
knet.get_next_vector(input_file_ptr);
// now apply it to the Kohonen network
knet.process_next_pattern();
dist_last_pattern=knet.get_win_dist();
// print result to output file

fprintf(output_file_ptr,”%i\t%i\t%i\t\t%i\t\t%f\n”,
total_cycles,total_patterns,knet.get_win_index(),
neighborhood_size,avg_dist_per_pattern);
total_patterns++;
// gradually reduce the neighborhood size

// and the gain, alpha
if (((total_cycles+1) % period) == 0)
{
if (neighborhood_size > 0)
neighborhood_size —;
knet.update_neigh_size(neighborhood_size);
if (alpha>0.1)
alpha −= (float)0.1;
}
dist_last_cycle += dist_last_pattern;
knet.update_weights(alpha);
dist_last_pattern = 0;
}
avg_dist_per_pattern= dist_last_cycle/patterns_per_cycle;
total_dist += dist_last_cycle;
total_cycles++;
fseek(input_file_ptr, 0L, SEEK_SET); // reset the file

pointer
// the file

} // end main loop
cout << “\n\n\n\n\n\n\n\n\n\n\n”;

cout << “—————————————————————————−\n”;
cout << “ done \n”;
avg_dist_per_cycle= total_dist/total_cycles;
cout << “\n”;

cout << “——>average dist per cycle = “ << avg_dist_per_cycle
<< “ <—−\n”;
cout << “——>dist last cycle = “ << dist_last_cycle << “ <—−
\n”;
cout << “−>dist last cycle per pattern= “ <<
avg_dist_per_pattern << “ <—−\n”;
cout << “——————>total cycles = “ << total_cycles << “ <—−\n”;
cout << “——————>total patterns = “ << total_patterns << “ <—−
\n”;
cout << “—————————————————————————−\n”;
// close the input file
fclose(input_file_ptr);
}


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Flow of the Program
The flow of the program is very similar to the backpropagation simulator. The criterion for ending the
simulation in the Kohonen program is the average winner distance. This is a Euclidean distance measure
between the input vector and the winner’s weight vector. This distance is the square root of the sum of the
squares of the differences between individual vector components between the two vectors.
Results from Running the Kohonen Program

Once you compile the program, you need to create an input file to try it. We will first use a very simple input
file and examine the results.
A Simple First Example
Let us create an input file, input.dat, which contains only two arbitrary vectors:
0.4 0.98 0.1 0.2

0.5 0.22 0.8 0.9
The file contains two four−dimensional vectors. We expect to see output that contains a
different winner neuron for each of these patterns. If this is the case, then the Kohonen map
has assigned different categories for each of the input vectors, and, in the future, you can
expect to get the same winner classification for vectors that are close to or equal to these
vectors.
By running the Kohonen map program, you will see the following output (user input is italic):
Please enter initial values for:

alpha (0.01−1.0),
and the neighborhood size (integer between 0 and 50)
separated by spaces, e.g. 0.3 5
0.3 5
Now enter the period, which is the
number of cycles after which the values
for alpha the neighborhood size are decremented
choose an integer between 1 and 500 , e.g. 50
50
Flow of the Program 237

500

A Kohonen network has an input layer
followed by a Kohonen (output) layer
4 10
——————————————————————————
done
——>average dist per cycle = 0.544275 <——−
——>dist last cycle = 0.0827523 <——−
−>dist last cycle per pattern= 0.0413762 <——−
——————>total cycles = 11 <——−
——————>total patterns = 22 <——−
——————————————————————————
The layer sizes are given as 4 for the input layer and 10 for the Kohonen layer. You should choose the size of
the Kohonen layer to be larger than the number of distinct patterns that you think are in the input data set. One
of the outputs reported on the screen is the distance for the last cycle per pattern. This value is listed as 0.04,
which is less than the terminating value set at the top of the kohonen.cpp file of 0.05. The map converged on a
solution. Let us look at the file, kohonen.dat, the output file, to see the mapping to winner indexes:
cycle pattern win index neigh_size avg_dist_per_pattern

——————————————————————————————————————————————————————————————————————
0 0 1 5 100.000000
0 1 3 5 100.000000
1 2 1 5 0.304285
1 3 3 5 0.304285
2 4 1 5 0.568255
2 5 3 5 0.568255
3 6 1 5 0.542793
3 7 8 5 0.542793
4 8 1 5 0.502416
4 9 8 5 0.502416
5 10 1 5 0.351692
5 11 8 5 0.351692
6 12 1 5 0.246184
6 13 8 5 0.246184
7 14 1 5 0.172329
7 15 8 5 0.172329
8 16 1 5 0.120630
8 17 8 5 0.120630
9 18 1 5 0.084441
9 19 8 5 0.084441
10 20 1 5 0.059109
10 21 8 5 0.059109
In this example, the neighborhood size stays at its initial value of 5. In the first column you see the cycle
number, and in the second the pattern number. Since there are two patterns per cycle, you see the cycle
number repeated twice for each cycle.
The Kohonen map was able to find two distinct winner neurons for each of the patterns. One
has winner index 1 and the other index 8.



by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Orthogonal Input Vectors Example
For a second example, look at Figure 11.5, where we choose input vectors on a two−dimensional unit circle
that are 90° apart. The input.dat file should look like the following:
1 0
0 1
−1 0
0 −1
Figure 11.5 Orthogonal input vectors.
Using the same parameters for the Kohonen network, but with layer sizes of 2 and 10, what result would you
expect? The output file, kohonen.dat, follows:

———————————————————————————————————————————————————————————————————————
0 0 4 5 100.000000
0 1 0 5 100.000000
0 2 9 5 100.000000
0 3 3 5 100.000000
1 4 4 5 0.444558
1 5 0 5 0.444558
497 1991 6 0 0.707107

498 1992 0 0 0.707107
498 1993 0 0 0.707107
498 1994 6 0 0.707107
498 1995 6 0 0.707107
499 1996 0 0 0.707107
499 1997 0 0 0.707107
499 1998 6 0 0.707107
499 1999 6 0 0.707107
You can see that this example doesn’t quite work. Even though the neighborhood size gradually got reduced
to zero, the four inputs did not get categorized to different outputs. The winner distance became stuck at the
value of 0.707, which is the distance from a vector at 45°. In other words, the map generalizes a little too
much, arriving at the middle value for all of the input vectors.
You can fix this problem by starting with a smaller neighborhood size, which provides for less generalization.
By using the same parameters and a neighborhood size of 2, the following output is obtained.
Orthogonal Input Vectors Example 240


—————————————————————————————————————————————————————————————————————
0 0 5 2 100.000000
0 1 6 2 100.000000
0 2 4 2 100.000000
0 3 9 2 100.000000
1 4 0 2 0.431695
1 5 6 2 0.431695
1 6 3 2 0.431695
1 7 9 2 0.431695
2 8 0 2 0.504728
2 9 6 2 0.504728
2 10 3 2 0.504728
2 11 9 2 0.504728
3 12 0 2 0.353309
3 13 6 2 0.353309
3 14 3 2 0.353309
3 15 9 2 0.353309
4 16 0 2 0.247317
4 17 6 2 0.247317
4 18 3 2 0.247317
4 19 9 2 0.247317
5 20 0 2 0.173122
5 21 6 2 0.173122
5 22 3 2 0.173122
5 23 9 2 0.173122
6 24 0 2 0.121185
6 25 6 2 0.121185
6 26 3 2 0.121185
6 27 9 2 0.121185
7 28 0 2 0.084830
7 29 6 2 0.084830
7 30 3 2 0.084830
7 31 9 2 0.084830
8 32 0 2 0.059381
8 33 6 2 0.059381
8 34 3 2 0.059381
8 35 9 2 0.059381
For this case, the network quickly converges on a unique winner for each of the four input patterns, and the
distance criterion is below the set criterion within eight cycles. You can experiment with other input data sets
and combinations of Kohonen network parameters.
Orthogonal Input Vectors Example 241


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Variations and Applications of Kohonen Networks

There are many variations of the Kohonen network. Some of these will be briefly discussed in this section.
Using a Conscience
DeSieno has used a conscience factor in a Kohonen network. For a winning neuron, if the neuron is winning
more than a fair share of the time (roughly more than 1/n, where n is the number of neurons), then this neuron
has a threshold that is applied temporarily to allow other neurons the chance to win. The purpose of this
modification is to allow more uniform weight distribution while learning is taking place.
LVQ: Learning Vector Quantizer
You have read about LVQ (Learning Vector Quantizer) in previous chapters. In light of the Kohonen map, it
should be pointed out that the LVQ is simply a supervised version of the Kohonen network. Inputs and
expected output categories are presented to the network for training. You get data clustered, just as a Kohonen
network, according to the similarity to other data inputs.
Counterpropagation Network
A neural network topology, called a counterpropagation network, is a combination of a Kohonen layer with a
Grossberg layer. This network was developed by Robert Hecht−Nielsen and is useful for prototyping of
systems, with a fairly rapid training time compared to backpropagation. The Kohonen layer provides for
categorization, while the Grossberg layer allows for Hebbian conditioned learning. Counterpropagation has
been used successfully in data compression applications for images. Compression ratios of 10:1 to 100:1 have
been obtained, using a lossy compression scheme that codes the image with a technique called vector
quantization, where the image is broken up into representative subimage vectors. The statistics of these
vectors is such that you find that a large part of the image can be adequately represented by a subset of all the
vectors. The vectors with the highest frequency of occurrence are coded with the shortest bit strings, hence
you achieve data compression.
Application to Speech Recognition
Kohonen created a phonetic typewriter by classifying speech waveforms of different phonemes of Finnish
speech into different categories using a Kohonen SOM. The Kohonen phoneme map used 50 samples of each
phoneme for calibration. These samples caused excitation in a neighborhood of cells more strongly than in
other cells. A neighborhood was labeled with the particular phoneme that caused excitation. For an utterance
of speech made to the network, the exact neighborhoods that were active during the utterance were noted, and
for how long, and in what sequence. Short excitations were taken as transitory sounds. The information
obtained from the network was then pieced together to find out the words in the utterance made to the
network.
Variations and Applications of Kohonen Networks 242

Summary
In this chapter, you have learned about one of the important types of competitive learning called Kohonen
feature map. The most significant points of this discussion are outlined as follows:
• The Kohonen feature map is an example of an unsupervised neural network that is mainly used as
a classifier system or data clustering system. As more inputs are presented to this network, the
network improves its learning and is able to adapt to changing inputs.
• The training law for the Kohonen network tries to align the weight vectors along the same
direction as input vectors.
• The Kohonen network models lateral competition as a form of self−organization. One winner
neuron is derived for each input pattern to categorize that input.
• Only neurons within a certain distance (neighborhood) from the winner are allowed to participate
in training for a given input pattern.
Summary 243

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Chapter 12
Application to Pattern Recognition
Using the Kohonen Feature Map
In this chapter, you will use the Kohonen program developed in Chapter 11 to recognize patterns. You will
modify the Kohonen program for the display of patterns.
An Example Problem: Character Recognition
The problem that is presented in this chapter is to recognize or categorize alphabetic characters. You will
input various alphabetic characters to a Kohonen map and train the network to recognize these as separate
categories. This program can be used to try other experiments that will be discussed at the end of this chapter.
Representing Characters
Each character is represented by a 5×7 grid of pixels. We use the graphical printing characters of the IBM
extended ASCII character set to show a grayscale output for each pixel. To represent the letter A, for example,
you could use the pattern shown in Figure 12.1. Here the blackened boxes represent value 1, while empty
boxes represent a zero. You can represent all characters this way, with a binary map of 35 pixel values.
Figure 12.1 Representation of the letter A with a 5×7 pattern.
The letter A is represented by the values:
0 0 1 0 0
0 1 0 1 0
1 0 0 0 1
1 0 0 0 1
1 1 1 1 1
1 0 0 0 1
1 0 0 0 1
For use in the Kohonen program, we need to serialize the rows, so that all entries appear on one line.
Chapter 12 Application to Pattern Recognition 244

For the characters A and X you would end up with the following entries in the input file, input.dat:
0 0 1 0 0 0 1 0 1 0 1 0 0 0 1 1 0 0 0 1 1 1 1 1 1 1 0 0 0 1 1 0 0 0 1
<< the letter A
1 0 0 0 1 0 1 0 1 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 1 0 1 0 1 0 0 0 1
<< the letter X
Monitoring the Weights

We will present the Kohonen map with many such characters and find the response in output. You will be
able to watch the Kohonen map as it goes through its cycles and learns the input patterns. At the same time,
you should be able to watch the weight vectors for the winner neurons to see the pattern that is developing in
the weights. Remember that for a Kohonen map the weight vectors tend to become aligned with the input
vectors. So after a while, you will notice that the weight vector for the input will resemble the input pattern
that you are categorizing.
Representing the Weight Vector

Although on and off values are fine for the input vectors mentioned, you need to see grayscale values for the
weight vector. This can be accomplished by quantizing the weight vector into four bins, each represented by a
different ASCII graphic character, as shown in Table 12.1.
Table 12.1 Quantizing the Weight Vector

<= 0White rectangle (space)
0 < weight <= 0.25Light−dotted rectangle
0.25 < weight <= 0.50Medium−dotted rectangle
0.50 < weight <= 0.75Dark−dotted rectangle
weight > 0.75Black rectangle
The ASCII values for the graphics characters to be used are listed in Table 12.2.
Table 12.2 ASCII Values for Rectangle Graphic Characters

White rectangle255
Light−dotted rectangle176
Medium−dotted rectangle177
Dark−dotted rectangle178
Black rectangle219
C++ Code Development

The changes to the Kohonen program are relatively minor. The following listing indicates these changes.
Changes to the Kohonen Program
The first change to make is to the Kohonen_network class definition. This is in the file, layerk.h, shown in
Listing 12.1.
Listing 12.1 Updated layerk.h file
class Kohonen_network
C++ Code Development 245

private:
layer *layer_ptr[2];
int layer_size[2];
public:
Kohonen_network();
~Kohonen_network();
void set_up_network(int);
void get_next_vector(FILE *);
void process_next_pattern();
int get_win_index();
void display_input_char();
void display_winner_weights();
};


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
The new member functions are shown in italic. The functions display_input_char() and
display_winner_weights() are used to display the input and weight maps on the screen to watch weight
character map converge to the input map.
The implementation of these functions is in the file, layerk.cpp. The portion of this file containing these
functions is shown in Listing 12.2.
Listing 12.2 Additions to the layerk.cpp implementation file
void Kohonen_network::display_input_char()
{
int i, num_inputs;
unsigned char ch;
float temp;
int col=0;
float * inputptr;
num_inputs=layer_ptr[1]−>num_inputs;
inputptr = layer_ptr[1]−>inputs;
// we’ve got a 5x7 character to display

{
temp = *(inputptr);
if (temp <= 0)
ch=255;// blank
else if ((temp > 0) && (temp <= 0.25))
ch=176; // dotted rectangle −light
else if ((temp > 0.25) && (temp <= 0.50))
ch=177; // dotted rectangle −medium
else if ((temp >0.50) && (temp <= 0.75))
ch=178; // dotted rectangle −dark
else if (temp > 0.75)
ch=219; // filled rectangle
printf(“%c”,ch); //fill a row
col++;
if ((col % 5)==0)
printf(“\n”); // new row
inputptr++;
}
printf(“\n\n\n”);
}
void Kohonen_network::display_winner_weights()
{
int i, k;
unsigned char ch;
float temp;
float * wmat;
int col=0;
int win_index;

int num_inputs, num_outputs;
num_inputs= layer_ptr[1]−>num_inputs;
wmat = ((Kohonen_layer*)layer_ptr[1])
−>weights;
win_index=((Kohonen_layer*)layer_ptr[1])
−>winner_index;
num_outputs=layer_ptr[1]−>num_outputs;
// we’ve got a 5x7 character to display

{
k= i*num_outputs;
temp = wmat[k+win_index];
if (temp <= 0)
ch=255;// blank
else if ((temp > 0) && (temp <= 0.25))
ch=176; // dotted rectangle −light
else if ((temp > 0.25) && (temp <= 0.50))
ch=177; // dotted rectangle −medium
else if ((temp > 0.50) && (temp <= 0.75))
ch=178; // dotted rectangle −dark
else if (temp > 0.75)
ch=219; // filled rectangle
printf(“%c”,ch); //fill a row
col++;
if ((col % 5)==0)
printf(“\n”); // new row
}
printf(“\n\n”);
printf(“—————————−\n”);
The final change to make is to the kohonen.cpp file. The new file is called pattern.cpp and is shown in Listing
12.3.
Listing 12.3 The implementation file pattern.cpp
// pattern.cpp V. Rao, H. Rao

// Kohonen map for pattern recognition
#include “layerk.cpp”
#define INPUT_FILE “input.dat”

#define OUTPUT_FILE “kohonen.dat”
#define dist_tol 0.001
#define wait_cycles 10000 // creates a pause to
// view the character maps
void main()
{
int neighborhood_size, period;

float avg_dist_per_cycle=0.0;
float dist_last_cycle=0.0;
float avg_dist_per_pattern=100.0; // for the latest cycle
float dist_last_pattern=0.0;
float total_dist;
float alpha;

unsigned startup;
int max_cycles;
int patterns_per_cycle=0;
int total_cycles, total_patterns;

int i;

Kohonen_network knet;
FILE * input_file_ptr, * output_file_ptr;
// open input file for reading

if ((input_file_ptr=fopen(INPUT_FILE,”r”))==NULL)
{
cout << “problem opening input file\n”;
exit(1);
}
// open writing file for writing

{
exit(1);
}
// ————————————————————−
// Read in an initial values for alpha, and the
// neighborhood size.
// Both of these parameters are decreased with
// time. The number of cycles to execute before
// decreasing the value of these parameters is
// called the period. Read in a value for the
// period.
// ————————————————————−
cout << “ Please enter initial values for:\n”;
cout << “alpha (0.01−1.0),\n”;
cout << “and the neighborhood size (integer between 0\
and 50)\n”;
cout << “separated by spaces, e.g. 0.3 5 \n “;
cin >> alpha >> neighborhood_size ;
cout << “\nNow enter the period, which is the\n”;

cout << “number of cycles after which the values\n”;
cout << “for alpha the neighborhood size are \
decremented\n”;
cout << “choose an integer between 1 and 500 , e.g. \ 50 \n”;
cin >> period;

cout << “\nPlease enter the maximum cycles for the
simulation\n”;
cout << “Try a value of 500 to start with\n\n”;
cin >> max_cycles;
// the main loop

//
// continue looping until the average distance is less than

// the tolerance specified at the top of this file

// , or the maximum number of
// cycles is exceeded;

knet.get_layer_info();

knet.set_up_network(neighborhood_size);
// randomize weights for the Kohonen layer

// note that the randomize function for the
// Kohonen simulator generates
// weights that are normalized to length = 1
knet.randomize_weights();
// write header to output file
“cycle\tpattern\twin index\tneigh_size\\
tavg_dist_per_pattern\n”);
“————————————————————————\n”);
startup=1;
total_dist=0;
while (
(avg_dist_per_pattern > dist_tol)
|| (startup==1)
)
{
startup=0;
dist_last_cycle=0; // reset for each cycle
while (!feof(input_file_ptr))
{
knet.get_next_vector(input_file_ptr);
// now apply it to the Kohonen network

knet.process_next_pattern();
dist_last_pattern=knet.get_win_dist();
// print result to output file

fprintf(output_file_ptr,”%i\t%i\t%i\t\t%i\t\t%f\n”,
total_cycles,total_patterns,knet.get_win_index(),
neighborhood_size,avg_dist_per_pattern);
// display the input character and the

// weights for the winner to see match
knet.display_input_char();

knet.display_winner_weights();
// pause for a while to view the
// character maps
for (i=0; i<wait_cycles; i++)
{;}
total_patterns++;
// gradually reduce the neighborhood size

// and the gain, alpha
if (((total_cycles+1) % period) == 0)
{
if (neighborhood_size > 0)
neighborhood_size —;
knet.update_neigh_size(neighborhood_size);
if (alpha>0.1)
alpha −= (float)0.1;
}
dist_last_cycle += dist_last_pattern;
knet.update_weights(alpha);
dist_last_pattern = 0;
}
avg_dist_per_pattern= dist_last_cycle/patterns_per_cycle;
total_dist += dist_last_cycle;
total_cycles++;
fseek(input_file_ptr, 0L, SEEK_SET); // reset the file

pointer
// the file
} // end main loop

cout << “———————————————————————\n”;
cout << “ done \n”;
avg_dist_per_cycle= total_dist/total_cycles;
cout << “\n”;

cout << “——>average dist per cycle = “ << avg_dist_per_cycle << “ <—−\n”;
cout << “>dist last cycle = “ << dist_last_cycle << “ < \n”;
cout << “−>dist last cycle per pattern= “ <<
avg_dist_per_pattern << “ <—−\n”;
cout << “——−>total cycles = “ << total_cycles << “ <—−\n”;
cout << “——————>total patterns = “ <<
total_patterns << “ <—−\n”;
cout << “————————————————————————\n”;
// close the input file
fclose(input_file_ptr);
}
Changes to the program are indicated in italic. Compile this program by compiling and making the pattern.cpp
file, after modifying the layerk.cpp and layerk.h files, as indicated previously.



by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Testing the Program
Let us run the example that we have created an input file for. We have an input.dat file with the characters A
and X defined. A run of the program with these inputs is shown as follows:
Please enter initial values for:

alpha (0.01−1.0),
and the neighborhood size (integer between 0 and 50)
separated by spaces, e.g., 0.3 5
0.3 5
Now enter the period, which is the
number of cycles after which the values
for alpha the neighborhood size are decremented
choose an integer between 1 and 500, e.g., 50
50
500
A Kohonen network has an input layer
followed by a Kohonen (output) layer
35 100
The output of the program is contained in file kohonen.dat as usual. This shows the following result.

———————————————————————————————————————————————————————————————————
0 0 42 5 100.000000
0 1 47 5 100.000000
1 2 42 5 0.508321
1 3 47 5 0.508321
2 4 40 5 0.742254
2 5 47 5 0.742254
3 6 40 5 0.560121
3 7 47 5 0.560121
4 8 40 5 0.392084
4 9 47 5 0.392084
5 10 40 5 0.274459
5 11 47 5 0.274459
6 12 40 5 0.192121
6 13 47 5 0.192121
7 14 40 5 0.134485
7 15 47 5 0.134485
8 16 40 5 0.094139
8 17 47 5 0.094139
9 18 40 5 0.065898
9 19 47 5 0.065898
10 20 40 5 0.046128
10 21 47 5 0.046128
Testing the Program 253

11 22 40 5 0.032290
11 23 47 5 0.032290
12 24 40 5 0.022603
12 25 47 5 0.022603
13 26 40 5 0.015822
13 27 47 5 0.015822
14 28 40 5 0.011075
14 29 47 5 0.011075
15 30 40 5 0.007753
15 31 47 5 0.007753
16 32 40 5 0.005427
16 33 47 5 0.005427
17 34 40 5 0.003799
17 35 47 5 0.003799
18 36 40 5 0.002659
18 37 47 5 0.002659
19 38 40 5 0.001861
19 39 47 5 0.001861
20 40 40 5 0.001303
20 41 47 5 0.001303
The tolerance for the distance was set to be 0.001 for this program, and the program was able to converge to
this value. Both of the inputs were successfully classified into two different winning output neurons. In
Figures 12.2 and 12.3 you see two snapshots of the input and weight vectors that you will find with this
program. The weight vector resembles the input as you can see, but it is not an exact replication.
Figure 12.2 Sample screen output of the letter A from the input and weight vectors.
Figure 12.3 Sample screen output of the letter X from the input and weight vectors.
Testing the Program 254


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
As mentioned in Chapter 11, you actually don’t desire the exact replication of the input pattern for the weight
vector. This would amount to memorizing of the input patterns with no capacity for generalization.
For example, a typical use of this alphabet classifier system would be to use it to process noisy data, like
handwritten characters. In such a case, you would need a great deal of latitude in scoping a class for a letter A.
Adding Characters
The next step of the program is to add characters and see what categories they end up in. There are many
alphabetic characters that look alike, such as H and B for example. You can expect the Kohonen classifier to
group these like characters into the same class.
We now modify the input.dat file to add the characters H, B, and I. The new input.dat file is shown as follows.
0 0 1 0 0 0 1 0 1 0 1 0 0 0 1 1 0 0 0 1 1 1 1 1 1 1 0 0 0 1
1 0 0 0 1
1 0 0 0 1 0 1 0 1 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 1 0 1 0
1 0 0 0 1
1 0 0 0 1 1 0 0 0 1 1 0 0 0 1 1 1 1 1 1 1 0 0 0 1 1 0 0 0 1
1 0 0 0 1
1 1 1 1 1 1 0 0 0 1 1 0 0 0 1 1 1 1 1 1 1 0 0 0 1 1 0 0 0 1
1 1 1 1 1
0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0
0 0 1 0 0
The output using this input file is shown as follows.
—————————————————————————−
done
——>average dist per cycle = 0.732607 <—−
——>dist last cycle = 0.00360096 <—−
−>dist last cycle per pattern= 0.000720192 <—−
——————>total cycles = 37 <—−
——————>total patterns = 185 <—−
—————————————————————————−
The file kohonen.dat with the output values is now shown as follows.

—————————————————————————————————————————————————————————————————
0 0 69 5 100.000000
0 1 93 5 100.000000
0 2 18 5 100.000000
0 3 18 5 100.000000
0 4 78 5 100.000000
Generalization versus Memorization 255

1 5 69 5 0.806743
1 6 93 5 0.806743
1 7 18 5 0.806743
1 8 18 5 0.806743
1 9 78 5 0.806743
2 10 69 5 0.669678
2 11 93 5 0.669678
2 12 18 5 0.669678
2 13 18 5 0.669678
2 14 78 5 0.669678
3 15 69 5 0.469631
3 16 93 5 0.469631
3 17 18 5 0.469631
3 18 18 5 0.469631
3 19 78 5 0.469631
4 20 69 5 0.354791
4 21 93 5 0.354791
4 22 18 5 0.354791
4 23 18 5 0.354791
4 24 78 5 0.354791
5 25 69 5 0.282990
5 26 93 5 0.282990
5 27 18 5 0.282990
...
35 179 78 5 0.001470
36 180 69 5 0.001029
36 181 93 5 0.001029
36 182 13 5 0.001029
36 183 19 5 0.001029
36 184 78 5 0.001029
Again, the network does not find a problem in classifying these vectors.
Until cycle 21, both the H and the B were classified as output neuron 18. The ability to
distinguish these vectors is largely due to the small tolerance we have assigned as a
termination criterion.
Generalization versus Memorization 256


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
You can try other experiments with the program. For example, you can repeat the input file but with the order
of the entries changed. In other words, you can present the same inputs a number of times in different order.
This actually helps the Kohonen network train faster. You can try applying garbled versions of the characters
to see if the network distinguishes them. Just as in the backpropagation program, you can save the weights in
a weight file to freeze the state of training, and then apply new inputs. You can enter all of the characters from
A to Z and see the classification that results. Do you need to train on all of the characters or a subset? You can
change the size of the Kohonen layer. How many neurons do you need to recognize the complete alphabet?
There is no restriction on using digital inputs of 1 and 0 as we had used. You can apply grayscale analog
values. The program will display the input pattern according to the quantization levels that were set. This set
can be expanded, and you can use a graphics interface to display more levels. You can then try pattern
recognition of arbitrary images, but remember that processing time will increase rapidly with the number of
neurons used. The number of input neurons you choose is dictated by the image resolution, unless you filter
and/or subsample the image before presenting it to the network. Filtering is the process of using a type of
averaging function applied to groups of pixels. Subsampling is the process of choosing a lower−output image
resolution by selecting fewer pixels than a source image. If you start with an image that is 100 × 100 pixels,
you can subsample this image 2:1 in each direction to obtain an image that is one−fourth the size, or 50 × 50
pixels. Whether you throw away every other pixel to get this output resolution or apply a filter is up to you.
You could average every two pixels to get one output pixel as an example of a very simple filter.
Summary
The following list highlights the important Kohonen program features which you learned in this chapter.
• This chapter presented a simple character recognition program using a Kohonen feature map.
• The input vectors and the weight vectors were displayed to show convergence and note similarity
between the two vectors.
• As training progresses, the weight vector for the winner neuron resembles the input character
map.
Other Experiments to Try 257


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Chapter 13
Backpropagation II
Enhancing the Simulator
In Chapter 7, you developed a backpropagation simulator. In this chapter, you will put it to use with examples
and also add some new features to the simulator: a term called momentum, and the capability of adding noise
to the inputs during simulation. There are many variations of the algorithm that try to alleviate two problems
with backpropagation. First, like other neural networks, there is a strong possibility that the solution found
with backpropagation is not a global error minimum, but a local one. You may need to shake the weights a
little by some means to get out of the local minimum, and possibly arrive at a lower minimum. The second
problem with backpropagation is speed. The algorithm is very slow at learning. There are many proposals for
speeding up the search process. Neural networks are inherently parallel processing architectures and are suited
for simulation on parallel processing hardware. While there are a few plug−in neural net or digital signal
processing boards available in the market, the low−cost simulation platform of choice remains the personal
computer. Speed enhancements to the training algorithm are therefore very buffernecessary.
Another Example of Using Backpropagation
Before modifying the simulator to add features, let’s look at the same problem we used the Kohonen map to
analyze in Chapter 12. As you recall, we would like to be able to distinguish alphabetic characters by
assigning them to different bins. For backpropagation, we would apply the inputs and train the network with
anticipated responses. Here is the input file that we used for distinguishing five different characters, A, X, H,
B, and I:
0 0 1 0 0 0 1 0 1 0 1 0 0 0 1 1 0 0 0 1 1 1 1 1 1 1 0 0 0 1
1 0 0 0 1 0 1 0 1 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 1 0 1 0
1 0 0 0 1 1 0 0 0 1 1 0 0 0 1 1 1 1 1 1 1 0 0 0 1 1 0 0 0 1
1 1 1 1 1 1 0 0 0 1 1 0 0 0 1 1 1 1 1 1 1 0 0 0 1 1 0 0 0 1
0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0
1 0 0 0 1
1 0 0 0 1
1 0 0 0 1
1 1 1 1 1
0 0 1 0 0
Each line has a 5x7 dot representation of each character. Now we need to name each of the output categories.
We can assign a simple 3−bit representation as follows:
A 000
Chapter 13 Backpropagation II 258

X 010
H 100
B 101
I 111
Let’s train the network to recognize these characters. The training.dat file looks like the following.
0 0 1 0 0 0 1 0 1 0 1 0 0 0 1 1 0 0 0 1 1 1 1 1 1 1 0 0 0 1
1 0 0 0 1 0 1 0 1 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 1 0 1 0
1 0 0 0 1 1 0 0 0 1 1 0 0 0 1 1 1 1 1 1 1 0 0 0 1 1 0 0 0 1
1 1 1 1 1 1 0 0 0 1 1 0 0 0 1 1 1 1 1 1 1 0 0 0 1 1 0 0 0 1
0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0
1 0 0 0 1 0 0 0
1 0 0 0 1 0 1 0
1 0 0 0 1 1 0 0
1 1 1 1 1 1 0 1
0 0 1 0 0 1 1 1
Now you can start the simulator. Using the parameters (beta = 0.1, tolerance = 0.001, and max_cycles = 1000)
and with three layers of size 35 (input), 5 (middle), and 3 (output), you will get a typical result like the
following.
−−−−−−−−−−−−−−−−−−−−−−−−−− −
−−>average error per cycle = 0.035713<−−

−−>error last cycle = 0.008223 <−−
−>error last cycle per pattern= 0.00164455 <−−
−−−−−−>total cycles = 1000 <−−
−−−−−−>total patterns = 5000 <−−
−−−−−−−−−−−−−−−−−−−−−−−−−−−
The simulator stopped at the 1000 maximum cycles specified in this case. Your results will be different since
the weights start at a random point. Note that the tolerance specified was nearly met. Let us see how close the
output came to what we wanted. Look at the output.dat file. You can see the match for the last pattern as
follows:
for input vector:

0.000000 0.000000 1.000000 0.000000 0.000000 0.000000 0.000000
1.000000 0.000000 0.000000 0.000000 0.000000 1.000000 0.000000
0.000000 0.000000 0.000000 1.000000 0.000000 0.000000 0.000000
0.000000 1.000000 0.000000 0.000000 0.000000 0.000000 1.000000
0.000000 0.000000 0.000000 0.000000 1.000000 0.000000 0.000000
output vector is:
0.999637 0.998721 0.999330
1.000000 1.000000 1.000000
−−−−−−−−−−−



by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
To see the outputs of all the patterns, we need to copy the training.dat file to the test.dat file and rerun the
simulator in Test mode. Remember to delete the expected output field once you copy the file.
Running the simulator in Test mode (0) shows the following result in the output.dat file:
for input vector:

0.000000 0.000000 1.000000 0.000000 0.000000 0.000000 1.000000
0.000000 1.000000 0.000000 1.000000 0.000000 0.000000 0.000000
1.000000 1.000000 0.000000 0.000000 0.000000 1.000000 1.000000
1.000000 1.000000 1.000000 1.000000 1.000000 0.000000 0.000000
0.000000 1.000000 1.000000 0.000000 0.000000 0.000000 1.000000
output vector is:
0.005010 0.002405 0.000141
−−−−−−−−−−−
for input vector:
1.000000 0.000000 0.000000 0.000000 1.000000 0.000000 1.000000
0.000000 1.000000 0.000000 0.000000 0.000000 1.000000 0.000000
0.000000 0.000000 0.000000 1.000000 0.000000 0.000000 0.000000
0.000000 1.000000 0.000000 0.000000 0.000000 1.000000 0.000000
1.000000 0.000000 1.000000 0.000000 0.000000 0.000000 1.000000
output vector is:
0.001230 0.997844 0.000663
−−−−−−−−−−−
for input vector:
1.000000 0.000000 0.000000 0.000000 1.000000 1.000000 0.000000
0.000000 0.000000 1.000000 1.000000 0.000000 0.000000 0.000000
1.000000 1.000000 1.000000 1.000000 1.000000 1.000000 1.000000
0.000000 0.000000 0.000000 1.000000 1.000000 0.000000 0.000000
0.000000 1.000000 1.000000 0.000000 0.000000 0.000000 1.000000
output vector is:
0.995348 0.000253 0.002677
−−−−−−−−−−−
for input vector:
1.000000 1.000000 1.000000 1.000000 1.000000 1.000000 0.000000
0.000000 0.000000 1.000000 1.000000 0.000000 0.000000 0.000000
1.000000 1.000000 1.000000 1.000000 1.000000 1.000000 1.000000
0.000000 0.000000 0.000000 1.000000 1.000000 0.000000 0.000000
0.000000 1.000000 1.000000 1.000000 1.000000 1.000000 1.000000
output vector is:
0.999966 0.000982 0.997594
−−−−−−−−−−−
for input vector:
0.000000 0.000000 1.000000 0.000000 0.000000 0.000000 0.000000
1.000000 0.000000 0.000000 0.000000 0.000000 1.000000 0.000000
0.000000 0.000000 0.000000 1.000000 0.000000 0.000000 0.000000
0.000000 1.000000 0.000000 0.000000 0.000000 0.000000 1.000000
0.000000 0.000000 0.000000 0.000000 1.000000 0.000000 0.000000
output vector is:
0.999637 0.998721 0.999330
−−−−−−−−−−−

The training patterns are learned very well. If a smaller tolerance is used, it would be possible to complete the
learning in fewer cycles. What happens if we present a foreign character to the network? Let us create a new
test.dat file with two entries for the letters M and J, as follows:
1 0 0 0 1 1 1 0 1 1 1 0 1 0 1 1 0 0 0 1 1 0 0 0 1 1 0 0 0 1
0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0
1 0 0 0 1
0 1 1 1 1
The results should show each foreign character in the category closest to it. The middle layer of the network
acts as a feature detector. Since we specified five neurons, we have given the network the freedom to define
five features in the input training set to use to categorize inputs. The results in the output.dat file are shown as
follows.
for input vector:

1.000000 0.000000 0.000000 0.000000 1.000000 1.000000 1.000000
0.000000 1.000000 1.000000 1.000000 0.000000 1.000000 0.000000
1.000000 1.000000 0.000000 0.000000 0.000000 1.000000 1.000000
0.000000 0.000000 0.000000 1.000000 1.000000 0.000000 0.000000
0.000000 1.000000 1.000000 0.000000 0.000000 0.000000 1.000000
output vector is:
0.963513 0.000800 0.001231
−−−−−−−−−−−
for input vector:
0.000000 0.000000 1.000000 0.000000 0.000000 0.000000 0.000000
1.000000 0.000000 0.000000 0.000000 0.000000 1.000000 0.000000
0.000000 0.000000 0.000000 1.000000 0.000000 0.000000 0.000000
0.000000 1.000000 0.000000 0.000000 0.000000 0.000000 1.000000
0.000000 0.000000 0.000000 1.000000 1.000000 1.000000 1.000000
output vector is:
0.999469 0.996339 0.999157
−−−−−−−−−−−
In the first pattern, an M is categorized as an H, whereas in the second pattern, a J is categorized as an I, as

expected. The case of the first pattern seems reasonable since the H and M share many pixels in common.

There are many other experiments you could try in order to get a better feel for how to train and use a
backpropagation neural network.
• You could use the ASCII 8−bit code to represent each character, and try to train the network. You
could also code all of the alphabetic characters and see if it’s possible to distinguish all of them.
• You can garble a character, to see if you still get the correct output.
• You could try changing the size of the middle layer, and see the effect on training time and
generalization ability.
• You could change the tolerance setting to see the difference between an overtrained and
undertrained network in generalization capability. That is, given a foreign pattern, is the network able
to find the closest match and use that particular category, or does it arrive at a new category
altogether?
We will return to the same example after enhancing the simulator with momentum and noise addition
capability.

Adding the Momentum Term

A simple change to the training law that sometimes results in much faster training is the addition of a
momentum term. The training law for backpropagation as implemented in the simulator is:
Weight change = Beta * output_error * input
Now we add a term to the weight change equation as follows:
Weight change = Beta * output_error * input +

Alpha*previous_weight_change
The second term in this equation is the momentum term. The weight change, in the absence of error, would be
a constant multiple by the previous weight change. In other words, the weight change continues in the
direction it was heading. The momentum term is an attempt to try to keep the weight change process moving,
and thereby not get stuck in local minimas.
Code Changes
The effected files to implement this change are the layer.cpp file, to modify the update_weights() member
function of the output_layer class, and the main backprop.cpp file to read in the value for alpha and pass it to
the member function. There is some additional storage needed for storing previous weight changes, and this
affects the layer.h file. The momentum term could be implemented in two ways:
1. Using the weight change for the previous pattern.

2. Using the weight change accumulated over the previous cycle.
Adding the Momentum Term 263


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Although both of these implementations are valid, the second is particularly useful, since it adds a term that is
significant for all patterns, and hence would contribute to global error reduction. We implement the second
choice by accumulating the value of the current cycle weight changes in a vector called cum_deltas. The past
cycle weight changes are stored in a vector called past_deltas. These are shown as follows in a portion of the
layer.h file.

{
protected:
float * weights;
float * cum_deltas; // for momentum
float * past_deltas; // for momentum
friend network;
...
Changes to the layer.cpp File

The implementation file for the layer class changes in the output_layer::update_weights() routine and the
constructor and destructor for output_layer. First, here is the constructor for output_layer. Changes are
highlighted in italic.
output_layer::output_layer(int ins, int outs)

{
int i, j, k;
num_inputs=ins;
num_outputs=outs;
cum_deltas = new float[num_inputs*num_outputs];
past_deltas = new float[num_inputs*num_outputs];
||(outputs==0)||(expected_values==0)
||(past_deltas==0)||(cum_deltas==0))
{
exit(1);
}
// zero cum_deltas and past_deltas matrix
{

k=i*num_outputs;
{
cum_deltas[k+j]=0;
past_deltas[k+j=0;
}
}
}
The destructor simply deletes the new vectors:
{
delete [num_outputs*num_inputs] past_deltas;
delete [num_outputs*num_inputs] cum_deltas;
}
Now let’s look at the update_weights() routine changes:
void output_layer::update_weights(const float beta,

const float alpha)
{
int i, j, k;
float delta;
// beta*output_error*input + alpha*past_delta
{
k=i*num_outputs;
{
delta=beta*output_errors[j]*(*(inputs+i))
+alpha*past_deltas[k+j];
weights[k+j] += delta;
cum_deltas[k+j]+=delta; // current cycle
}
The change to the training law amounts to calculating a delta and adding it to the cumulative total of weight
changes in cum_deltas. At some point (at the start of a new cycle) you need to set the past_deltas vector to
the cum_delta vector. Where does this occur? Since the layer has no concept of cycle, this must be done at
the network level. There is a network level function called update_momentum at the beginning of each cycle
that in turns calls a layer level function of the same name. The layer level function swaps the past_deltas
vector and the cum_deltas vector, and reinitializes the cum_deltas vector to zero. We need to return to the
layer.h file to see changes that are needed to define the two functions mentioned.

{
protected:

float * weights;
friend network;
public:
~output_layer();
void update_weights(const float, const float);
void update_momentum();
void list_errors();
};
class network
private:
float *buffer;
fpos_t position;
unsigned training;
public:
network();
~network();
...
At both the network and output_layer class levels the function prototype for the update_momentum
member functions are highlighted. The implementation for these functions are shown as follows from the
layer.cpp class.
void output_layer::update_momentum()
{
// This function is called when a
// new cycle begins; the past_deltas
// pointer is swapped with the
// cum_deltas pointer. Then the contents
// pointed to by the cum_deltas pointer

// is zeroed out.
int i, j, k;
float * temp;
// swap
temp = past_deltas;
past_deltas=cum_deltas;
cum_deltas=temp;
// zero cum_deltas matrix

// for new cycle
{
k=i*num_outputs;
cum_deltas[k+j]=0;
}
}
void network::update_momentum()
{
int i;

−>update_momentum();
}


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Adding Noise During Training

Another approach to breaking out of local minima as well as to enhance generalization ability is to introduce
some noise in the inputs during training. A random number is added to each input component of the input
vector as it is applied to the network. This is scaled by an overall noise factor, NF, which has a 0 to 1 range.
You can add as much noise to the simulation as you want, or not any at all, by choosing NF = 0. When you
are close to a solution and have reached a satisfactory minimum, you don’t want noise at that time to interfere
with convergence to the minimum. We implement a noise factor that decreases with the number of cycles, as
shown in the following excerpt from the backprop.cpp file.
// update NF
// gradually reduce noise to zero
if (total_cycles>0.7*max_cycles)
new_NF = 0;
else if (total_cycles>0.5*max_cycles)
new_NF = 0.25*NF;
new_NF = 0.50*NF;
new_NF = 0.75*NF;
backp.set_NF(new_NF);
The noise factor is reduced at regular intervals. The new noise factor is updated with the network class
function called set_NF(float). There is a member variable in the network class called NF that holds the
current value for the noise factor. The noise is added to the inputs in the input_layer member function
calc_out().
Another reason for using noise is to prevent memorization by the network. You are effectively presenting a
different input pattern with each cycle so it becomes hard for the network to memorize patterns.
One Other Change—Starting Training from a Saved Weight File

Shortly, we will look at the complete listings for the backpropagation simulator. There is one other
enhancement to discuss. It is often useful in long simulations to be able to start from a known point, which is
from an already saved set of weights. This is a simple change in the backprop.cpp program, which is well
worth the effort. As a side benefit, this feature will allow you to run a simulation with a large beta value for,
say, 500 cycles, save the weights, and then start a new simulation with a smaller beta value for another 500 or
more cycles. You can take preset breaks in long simulations, which you will encounter in Chapter 14. At this
point, let’s look at the complete listings for the updated layer.h and layer.cpp files in Listings 13.1 and 13.2:
Listing 13.1 layer.h file updated to include noise and momentum
Adding Noise During Training 268


// added noise and momentum
class network;
class layer
{
protected:
int num_inputs;
int num_outputs;
float *outputs; // pointer to array of outputs
friend network;
friend Kohonen_network; // update for Kohonen model
public:

};
class input_layer: public layer

{
private:
float noise_factor;
float * orig_outputs;
public:
input_layer(int, int);
~input_layer();
void set_NF(float);
friend network;
};
class middle_layer;

{
protected:
float * weights;
friend network;

public:
~output_layer();
void list_errors();
};
class middle_layer: public output_layer

{
private:
public:
middle_layer(int, int);
~middle_layer();
void calc_error();
};
class network
private:
float *buffer;
fpos_t position;
unsigned training;
public:
network();
~network();
void write_weights(FILE *);
void read_weights(FILE *);
void write_outputs(FILE *);
void list_errors();
void forward_prop();
void backward_prop(float &);
int fill_IObuffer(FILE *);
void set_up_pattern(int);
void set_NF(float);

};


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Listing 13.2 Layer.cpp file updated to include noise and momentum
// layer.cpp V.Rao, H.Rao

// added momentum and noise

#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <time.h>
#include "layer.h"
inline float squash(float input)

// squashing function
// use sigmoid — can customize to something
// else if desired; can add a bias term too
//
{
if (input < −50)
return 0.0;
else if (input > 50)
return 1.0;
else return (float)(1/(1+exp(−(double)input)));
inline float randomweight(unsigned init)

{
int num;
// will return a floating point
// value between −1 and 1
if (init==1) // seed the generator

num=rand() % 100;
return 2*(float(num/100.00))−1;
}
// the next function is needed for Turbo C++

// and Borland C++ to link in the appropriate
// functions for fscanf floating point formats:
static void force_fpf()
{
float x, *y;
y=&x;
x=*y;
}

// −−−−−−−−−−−−−−−−−−−−−
// input layer
//−−−−−−−−−−−−−−−−−−−−−
input_layer::input_layer(int i, int o)
{
num_inputs=i;
num_outputs=o;

orig_outputs = new float[num_outputs];
if ((outputs==0)||(orig_outputs==0))
{
exit(1);
}
noise_factor=0;
input_layer::~input_layer()
{
delete [num_outputs] orig_outputs;
}
void input_layer::calc_out()
{
//add noise to inputs
// randomweight returns a random number
// between −1 and 1
int i;
for (i=0; i<num_outputs; i++)
outputs[i] =orig_outputs[i]*
(1+noise_factor*randomweight(0));
void input_layer::set_NF(float noise_fact)

{
noise_factor=noise_fact;
}
// −−−−−−−−−−−−−−−−−−−−−
// output layer
//−−−−−−−−−−−−−−−−−−−−−
output_layer::output_layer(int ins, int outs)

{
int i, j, k;
num_inputs=ins;
num_outputs=outs;
cum_deltas = new float[num_inputs*num_outputs];
past_deltas = new float[num_inputs*num_outputs];

||(outputs==0)||(expected_values==0)
||(past_deltas==0)||(cum_deltas==0))
{
exit(1);
}
// zero cum_deltas and past_deltas matrix

{
k=i*num_outputs;
{
cum_deltas[k+j]=0;
past_deltas[k+j]=0;
}
}
}
{
delete [num_outputs*num_inputs] past_deltas;
delete [num_outputs*num_inputs] cum_deltas;
void output_layer::calc_out()
{
int i,j,k;

{
{
k=i*num_outputs;
{
cout << "weights are blowing up\n";
cout << "try a smaller learning constant\n";
cout << "e.g. beta=0.02 aborting...\n";
exit(1);
}
}
// use the sigmoid squash function
outputs[j]=squash(accumulator);
accumulator=0;
}

void output_layer::calc_error(float & error)

{
int i, j, k;
float total_error=0;
{
output_errors[j] = expected_values[j]−outputs[j];
total_error+=output_errors[j];
}
error=total_error;

{
k=i*num_outputs;
{
back_errors[i]=
weights[k+j]*output_errors[j];
}
accumulator=0;
}
void output_layer::randomize_weights()
{
int i, j, k;

float discard;

{
k=i*num_outputs;
}
}
void output_layer::update_weights(const float beta,

const float alpha)
{
int i, j, k;
float delta;

// beta*output_error*input + alpha*past_delta


{
k=i*num_outputs;
{
delta=beta*output_errors[j]*(*(inputs+i))
+alpha*past_deltas[k+j];
weights[k+j] += delta;
cum_deltas[k+j]+=delta; // current cycle
}
void output_layer::update_momentum()
{
// This function is called when a
// new cycle begins; the past_deltas
// pointer is swapped with the
// cum_deltas pointer. Then the contents
// pointed to by the cum_deltas pointer
// is zeroed out.
int i, j, k;
float * temp;
// swap
temp = past_deltas;
past_deltas=cum_deltas;
cum_deltas=temp;
// zero cum_deltas matrix

// for new cycle
{
k=i*num_outputs;
cum_deltas[k+j]=0;
}
}
void output_layer::list_weights()
{
int i, j, k;

{
k=i*num_outputs;
cout << "weight["<<i<<","<<
j<<"] is: "<<weights[k+j];
}
void output_layer::list_errors()
{
int i, j;

cout << "backerror["<<i<<
"] is : "<<back_errors[i]<<"\n";


cout << "outputerrors["<<j<<
"] is: "<<output_errors[j]<<"\n";
void output_layer::write_weights(int layer_no,

{
int i, j, k;

// writing
// prepend the layer_no to all lines of data

// format:
// ...

{
fprintf(weights_file_ptr,"%i ",layer_no);
k=i*num_outputs;
{
fprintf(weights_file_ptr,"%f ",
weights[k+j]);
}
fprintf(weights_file_ptr,"\n");
}
void output_layer::read_weights(int layer_no,

{
int i, j, k;

// reading
// look for the prepended layer_no

// format:
// ...
while (1)
fscanf(weights_file_ptr,"%i";,&j);
if ((j==layer_no)|| (feof(weights_file_ptr)))
break;
else
{
while (fgetc(weights_file_ptr) != `\n')
{;}// get rest of line
}
}
if (!(feof(weights_file_ptr)))

{
// continue getting first line
i=0;
{
&weights[j]); // i*num_outputs
= 0
}
// now get the other lines

{
fscanf(weights_file_ptr,
”%i”,&layer_no);
k=i*num_outputs;
{
fscanf(weights_file_ptr,”%f”,
&weights[k+j]);
}
}
fscanf(weights_file_ptr,”\n”);
}
else cout << “end of file reached\n”;
void output_layer::list_outputs()
{
int j;

{
cout << “outputs[“<<j
<<”] is: “<<outputs[j]<<”\n”;
}
// ————————————————————−
// middle layer
//—————————————————————
middle_layer::middle_layer(int i, int o):

output_layer(i,o)
{
}
middle_layer::~middle_layer()
{
}
void middle_layer::calc_error()

{
int i, j, k;

{
k=i*num_outputs;
{
back_errors[i]=
weights[k+j]*(*(output_errors+j));
}
accumulator=0;
}
network::network()
{
position=0L;
}
network::~network()
{
int i,j,k;
k=MAX_VECTORS;
delete [(i+j)*k]buffer;
}
void network::set_training(const unsigned & value)

{
training=value;
}
unsigned network::get_training_value()
{
return training;
}
void network::get_layer_info()
{
int i;
//—————————————————————
//
// Get layer sizes for the network
//
// ————————————————————−
cout << “ Please enter in the number of layers for your net work.\n”;
cout << “ You can have a minimum of 3 to a maximum of 5. \n”;
cout << “ 3 implies 1 hidden layer; 5 implies 3 hidden layers : \n\n”;

cin >> number_of_layers;
cout << “ Enter in the layer sizes separated by spaces.\n”;

cout << “ For a network with 3 neurons in the input layer,\n”;
cout << “ 2 neurons in a hidden layer, and 4 neurons in the\n”;
cout << “ output layer, you would enter: 3 2 4 .\n”;
cout << “ You can have up to 3 hidden layers,for five maximum entries
:\n\n”;

{
}
// ———————————————————————————
// size of layers:
// output_layer layer_size[number_of_layers−1]
// middle_layers layer_size[1]
//———————————————————————————−
void network::set_up_network()
{
int i,j,k;
//———————————————————————————−
//
//———————————————————————————−

{
layer_ptr[i+1] =
new middle_layer(layer_size[i],layer_size[i+1]);
}
layer_ptr[number_of_layers−1] = new
output_layer(layer_size[number_of_layers−2], layer_size[number_of_
layers−1]);
{
{
cout << “insufficient memory\n”;
cout << “use a smaller architecture\n”;
exit(1);
}
}
//———————————————————————————−
//
//———————————————————————————−
// set inputs to previous layer outputs for all layers,
// except the input layer
for (i=1; i< number_of_layers; i++)

layer_ptr[i]−>inputs = layer_ptr[i−1]−>outputs;
// for back_propagation, set output_errors to next layer

// back_errors for all layers except the output
// layer and input layer
for (i=1; i< number_of_layers −1; i++)

((output_layer *)layer_ptr[i])−>output_errors =
((output_layer *)layer_ptr[i+1])−>back_errors;
// define the IObuffer that caches data from

// the datafile
k=MAX_VECTORS;
buffer=new
float[(i+j)*k];
if (buffer==0)
{
cout << “insufficient memory for buffer\n”;
exit(1);
}
}
void network::randomize_weights()
{
int i;

}
void network::update_weights(const float beta, const float alpha)

{
int i;

−>update_weights(beta,alpha);
}
void network::update_momentum()
{
int i;
−>update_momentum();
}
void network::write_weights(FILE * weights_file_ptr)

{
int i;

−>write_weights(i,weights_file_ptr);
}
void network::read_weights(FILE * weights_file_ptr)

{

int i;

−>read_weights(i,weights_file_ptr);
}
void network::list_weights()
{
int i;

{
cout << “layer number : “ <<i<< “\n”;
−>list_weights();
}
}
void network::list_outputs()
{
int i;

{
−>list_outputs();
}
}
void network::write_outputs(FILE *outfile)

{
int i, ins, outs;
float temp;
fprintf(outfile,”for input vector:\n”);

{
temp=layer_ptr[0]−>outputs[i];
fprintf(outfile,”%f “,temp);
}
fprintf(outfile,”\noutput vector is:\n”);

{
temp=layer_ptr[number_of_layers−1]−>
outputs[i];
if (training==1)
{
fprintf(outfile,”\nexpected output vector is:\n”);

{

temp=((output_layer *)(layer_ptr[number_of_layers−1]))−>
expected_values[i];
}
}
fprintf(outfile,”\n———————————\n”);
void network::list_errors()
{
int i;

{
−>list_errors();
}
}
int network::fill_IObuffer(FILE * inputfile)

{
// this routine fills memory with
// an array of input, output vectors
// up to a maximum capacity of
// MAX_INPUT_VECTORS_IN_ARRAY
// the return value is the number of read
// vectors
int i, k, count, veclength;
int ins, outs;
if (training==1)
veclength=ins+outs;
else
veclength=ins;
count=0;
while ((count<MAX_VECTORS)&&
(!feof(inputfile)))
{
k=count*(veclength);
for (i=0; i<veclength; i++)
{
fscanf(inputfile,”%f”,&buffer[k+i]);
}
fscanf(inputfile,”\n”);
count++;
}
if (!(ferror(inputfile)))
return count;
else return −1; // error condition

void network::set_up_pattern(int buffer_index)

{
// read one vector into the network
int i, k;
int ins, outs;
if (training==1)
k=buffer_index*(ins+outs);
else
k=buffer_index*ins;

((input_layer*)layer_ptr[0])
−>orig_outputs[i]=buffer[k+i];
if (training==1)
{
((output_layer *)layer_ptr[number_of_layers−1])−>
expected_values[i]=buffer[k+i+ins];
}
void network::forward_prop()
{
int i;
{
layer_ptr[i]−>calc_out(); //polymorphic
// function
}
}
void network::backward_prop(float & toterror)

{
int i;
// error for the output layer
((output_layer*)layer_ptr[number_of_layers−1])−>
calc_error(toterror);
// error for the middle layer(s)

for (i=number_of_layers−2; i>0; i—)
{
((middle_layer*)layer_ptr[i])−>
calc_error();
void network::set_NF(float noise_fact)

{
((input_layer*)layer_ptr[0])−>set_NF(noise_fact);
}



by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
The New and Final backprop.cpp File

The last file to present is the backprop.cpp file. This is shown in Listing 13.3.
Listing 13.3 Implementation file for the backpropagation simulator, with noise and momentum backprop.cpp
// backprop.cpp V. Rao, H. Rao

#include “layer.cpp”
#define TRAINING_FILE “training.dat”

#define WEIGHTS_FILE “weights.dat”
#define OUTPUT_FILE “output.dat”
#define TEST_FILE “test.dat”
void main()
{
float error_tolerance=0.1;
float total_error=0.0;
float avg_error_per_cycle=0.0;
float error_last_cycle=0.0;
float avgerr_per_pattern=0.0; // for the latest cycle
float error_last_pattern=0.0;
float learning_parameter=0.02;
float alpha; // momentum parameter
float NF; // noise factor
float new_NF;
unsigned temp, startup, start_weights;

long int vectors_in_buffer;
long int max_cycles;
long int patterns_per_cycle=0;
long int total_cycles, total_patterns;

int i;

network backp;
FILE * training_file_ptr, * weights_file_ptr, * output_file_ptr;

FILE * test_file_ptr, * data_file_ptr;
// open output file for writing

{
exit(1);
}
// enter the training mode : 1=training on 0=training off

cout << “—————————————————————————−\n”;

cout << “ C++ Neural Networks and Fuzzy Logic \n”;
cout << “ Backpropagation simulator \n”;
cout << “ version 2 \n”;
cout << “—————————————————————————−\n”;
cout << “Please enter 1 for TRAINING on, or 0 for off: \n\n”;
cout << “Use training to change weights according to your\n”;
cout << “expected outputs. Your training.dat file should contain\n”;
cout << “a set of inputs and expected outputs. The number of\n”;
cout << “inputs determines the size of the first (input) layer\n”;
cout << “while the number of outputs determines the size of the\n”;
cout << “last (output) layer :\n\n”;
cin >> temp;

backp.set_training(temp);
if (backp.get_training_value() == 1)
{
cout << “—> Training mode is *ON*. weights will be saved\n”;
cout << “in the file weights.dat at the end of the\n”;
cout << “current set of input (training) data\n”;
}
else
{
cout << “—> Training mode is *OFF*. weights will be loaded\n”;
cout << “from the file weights.dat and the current\n”;
cout << “(test) data set will be used. For the test\n”;
cout << “data set, the test.dat file should contain\n”;
cout << “only inputs, and no expected outputs.\n”;
}
{
// ————————————————————−
// Read in values for the error_tolerance,
// and the learning_parameter
// ————————————————————−
cout << “ Please enter in the error_tolerance\n”;
cout << “ —− between 0.001 to 100.0, try 0.1 to start − \n”;
cout << “\n”;
cout << “and the learning_parameter, beta\n”;
cout << “ —− between 0.01 to 1.0, try 0.5 to start − \n\n”;
cout << “ separate entries by a space\n”;
cout << “ example: 0.1 0.5 sets defaults mentioned :\n\n”;
cin >> error_tolerance >> learning_parameter;
// ————————————————————−
// Read in values for the momentum
// parameter, alpha (0−1.0)
// and the noise factor, NF (0−1.0)
// ————————————————————−
cout << “Enter values now for the momentum \n”;
cout << “parameter, alpha(0−1.0)\n”;
cout << “ and the noise factor, NF (0−1.0)\n”;
cout << “You may enter zero for either of these\n”;
cout << “parameters, to turn off the momentum or\n”;
cout << “noise features.\n”;
cout << “If the noise feature is used, a random\n”;
cout << “component of noise is added to the inputs\n”;
cout << “This is decreased to 0 over the maximum\n”;
cout << “number of cycles specified.\n”;
cout << “enter alpha followed by NF, e.g., 0.3 0.5\n”;

cin >> alpha >> NF;
//—————————————————————−
// open training file for reading
//—————————————————————−
if ((training_file_ptr=fopen(TRAINING_FILE,”r”))==NULL)
{
cout << “problem opening training file\n”;
exit(1);
}
data_file_ptr=training_file_ptr; // training on

cout << “Please enter the maximum cycles for the simulation\n”;
cout << “Try a value of 10 to start with\n”;
cin >> max_cycles;

cout << “Do you want to read weights from weights.dat to
start?\n”;
cout << “Type 1 to read from file, 0 to randomize starting
weights\n”;
cin >> start_weights;
}
else
{
if ((test_file_ptr=fopen(TEST_FILE,”r”))==NULL)
{
cout << “problem opening test file\n”;
exit(1);
}
data_file_ptr=test_file_ptr; // training off

}
// training: continue looping until the total error is less than

// the tolerance specified, or the maximum number of
// cycles is exceeded; use both the forward signal
propagation
// and the backward error propagation phases. If the error
// tolerance criteria is satisfied, save the weights in a
file.
// no training: just proceed through the input data set once in the
// forward signal propagation phase only. Read the starting
// weights from a file.
// in both cases report the outputs on the screen
new_NF=NF;

backp.get_layer_info();

backp.set_up_network();

if ((backp.get_training_value()==1)&&(start_weights!=1))
{
// randomize weights for all layers; there is no
// weight matrix associated with the input layer
// weight file will be written after processing
backp.randomize_weights();
// set up the noise factor value
}
else
{
// read in the weight matrix defined by a
// prior run of the backpropagation simulator
// with training on
if ((weights_file_ptr=fopen(WEIGHTS_FILE,”r”))
==NULL)
{
cout << “problem opening weights file\n”;
exit(1);
}
backp.read_weights(weights_file_ptr);
}
// main loop
// if training is on, keep going through the input data
// until the error is acceptable or the maximum number of
cycles
// is exceeded.
// if training is off, go through the input data once. report outputs
// with inputs to file output.dat
startup=1;
vectors_in_buffer = MAX_VECTORS; // startup condition
total_error = 0;
while ( ((backp.get_training_value()==1)
&& (avgerr_per_pattern
> error_tolerance)
&& (vectors_in_buffer !=0))
&& (total_cycles < 1))
&& (startup==1))
)
{
startup=0;
error_last_cycle=0; // reset for each cycle
backp.update_momentum(); // added to reset

// momentum matrices
// each cycle

// going through one buffer at a time
// pattern by pattern
while ((vectors_in_buffer==MAX_VECTORS))
{

vectors_in_buffer=
backp.fill_IObuffer(data_file_ptr); // fill buffer
if (vectors_in_buffer < 0)
{
cout << “error in reading in vectors, aborting\n”;
cout << “check that there are no extra linefeeds\n”;
cout << “in your data file, and that the number\n”;
cout << “of layers and size of layers match the\n”;
cout << “the parameters provided.\n”;
exit(1);
}
// process vectors
for (i=0; i<vectors_in_buffer; i++)
{
// get next pattern
backp.set_up_pattern(i);
total_patterns++;
// forward propagate
backp.forward_prop();
// back_propagate, if appropriate
{
backp.backward_prop(error_last_pattern);
error_last_cycle +=
error_last_pattern*error_last_pattern;
avgerr_per_pattern=
((float)sqrt((double)error_last_cycle/patterns_per_cycle));
// if it’s not the last cycle, update weights
if ((avgerr_per_pattern
> error_tolerance)
&& (total_cycles+1 < max_cycles))
backp.update_weights(learning_
parameter, alpha);
// backp.list_weights(); // can
// see change in weights by
// using list_weights before and
// after back_propagation
}
}
error_last_pattern = 0;
}
total_error += error_last_cycle;
total_cycles++;
// update NF
// gradually reduce noise to zero
if (total_cycles>0.7*max_cycles)
new_NF = 0;

new_NF = 0.25*NF;
new_NF = 0.50*NF;
new_NF = 0.75*NF;
// most character displays are 25 lines

// user will see a corner display of the cycle count
// as it changes
cout << “\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n”;

cout << total_cycles << “\t” << avgerr_per_pattern << “\n”;
fseek(data_file_ptr, 0L, SEEK_SET); // reset the file pointer

// the file
vectors_in_buffer = MAX_VECTORS; // reset
} // end main loop
{
if ((weights_file_ptr=fopen(WEIGHTS_FILE,”w”))
==NULL)
{
cout << “problem opening weights file\n”;
exit(1);
}
}

cout << “————————————————————————\n”;
cout << “ done: results in file output.dat\n”;
cout << “ training: last vector only\n”;
cout << “ not training: full cycle\n\n”;
{
backp.write_weights(weights_file_ptr);
avg_error_per_cycle=(float)sqrt((double)total_error/
total_cycles);
error_last_cycle=(float)sqrt((double)error_last_cycle);
cout << “ weights saved in file

weights.dat\n”;
cout << “\n”;
cout << “——>average error per cycle =
“ << avg_error_per_cycle << “
<—−\n”;
cout << “——>error last cycle = “
<< error_last_cycle << “ <—−\n”;
???cout << “−>error last cycle per
pattern=“<<avgerr_per_pattern
<<“<—−\n”;
cout << “——————>total

cycles = “ << total_cycles << “

<—−\n”;
cout << “——————>total
patterns = “ << total_patterns
<< “ <—−\n”;
cout <<
“——————————
;——————————————\n”;
// close all files
fclose(data_file_ptr);
fclose(output_file_ptr);


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Trying the Noise and Momentum Features

You can test out the version 2 simulator, which you just compiled with the example that you saw at the
beginning of the chapter. You will find that there is a lot of trial and error in finding optimum values for
alpha, the noise factor, and beta. This is true also for the middle layer size and the number of middle layers.
For some problems, the addition of momentum makes convergence much faster. For other problems, you may
not find any noticeable difference. An example run of the five−character recognition problem discussed at the
beginning of this chapter resulted in the following results with beta = 0.1, tolerance = 0.001, alpha = 0.25,
NF = 0.1, and the layer sizes kept at 35 5 3.
—————————————————————————−
——>average error per cycle = 0.02993<—−

——>error last cycle = 0.00498<—−
−>error last cycle per pattern= 0.000996 <—−
——————>total cycles = 242 <—−
——————>total patterns = 1210 <—−
—————————————————————————−
The network was able to converge on a better solution (in terms of error measurement) in
one−fourth the number of cycles. You can try varying alpha and NF to see the effect on
overall simulation time. You can now start from the same initial starting weights by
specifying a value of 1 for the starting weights question. For large values of alpha and beta,
the network usually will not converge, and the weights will get unacceptably large (you will
receive a message to that effect).
Variations of the Backpropagation Algorithm

Backpropagation is a versatile neural network algorithm that very often leads to success. Its Achilles heel is
the slowness at which it converges for certain problems. Many variations of the algorithm exist in the
literature to try to improve convergence speed and robustness. Variations have been proposed in the following
portions of the algorithm:
• Adaptive parameters. You can set rules that modify alpha, the momentum parameter, and beta,
the learning parameter, as the simulation progresses. For example, you can reduce beta whenever a
weight change does not reduce the error. You can consider undoing the particular weight change,
setting alpha to zero and redoing the weight change with the new value of beta.
Trying the Noise and Momentum Features 293

• Use other minimum search routines besides steepest descent. For example, you could use
Newton’s method for finding a minimum, although this would be a fairly slow process. Other
examples include the use of conjugate gradient methods or Levenberg−Marquardt optimization, both
of which would result in very rapid training.
• Use different cost functions. Instead of calculating the error (as expected—actual output), you
could determine another cost function that you want to minimize.
• Modify the architecture. You could use partially connected layers instead of fully connected
layers. Also, you can use a recurrent network, that is, one in which some outputs feed back as inputs.
Applications
Backpropagation remains the king of neural network architectures because of its ease of use and wide
applicability. A few of the notable applications in the literature will be cited as examples.
• NETTalk. In 1987, Sejnowski and Rosenberg developed a network connected to a speech

synthesizer that was able to utter English words, being trained to produce phonemes from English
text. The architecture consisted of an input layer window of seven characters. The characters were
part of English text that was scrolled by. The network was trained to pronounce the letter at the center
of the window. The middle layer had 80 neurons, while the output layer consisted of 26 neurons. With
1024 training patterns and 10 cycles, the network started making intelligible speech, similar to the
process of a child learning to talk. After 50 cycles, the network was about 95% accurate. You could
purposely damage the network with the removal of neurons, but this did not cause performance to
drop off a cliff; instead, the performance degraded gracefully. There was rapid recovery with
retraining using fewer neurons also. This shows the fault tolerance of neural networks.
• Sonar target recognition. Neural nets using backpropagation have been used to identify different
types of targets using the frequency signature (with a Fast Fourier transform) of the reflected signal.
• Car navigation. Pomerleau developed a neural network that is able to navigate a car based on
images obtained from a camera mounted on the car’s roof, and a range finder that coded distances in
grayscale. The 30×32 pixel image and the 8×32 range finder image were fed into a hidden layer of
size 29 feeding an output layer of 45 neurons. The output neurons were arranged in a straight line
with each side representing a turn to a particular direction (right or left), while the center neurons
represented “drive straight ahead.” After 1200 road images were trained on the network, the neural
network driver was able to negotiate a part of the Carnegie−Mellon campus at a speed of about 3
miles per hour, limited only by the speed of the real−time calculations done on a trained network in
the Sun−3 computer in the car.
• Image compression. G.W. Cottrell, P. Munro, and D. Zipser used backpropagation to compress
images with the result of an 8:1 compression ratio. They used standard backpropagation with 64 input
neurons (8×8 pixels), 16 hidden neurons, and 64 output neurons equal to the inputs. This is called
self−supervised backpropagation and represents an autoassociative network. The compressed signal
is taken from the hidden layer. The input to hidden layer comprised the compressor, while the hidden
to output layer forms a decompressor.
• Image recognition. Le Cun reported a backpropagation network with three hidden layers that
could recognize handwritten postal zip codes. He used a 16×16 array of pixel to represent each
handwritten digit and needed to encode 10 outputs, each of which represented a digit from 0 to 9. One
interesting aspect of this work is that the hidden layers were not fully connected. The network was set
up with blocks of neurons in the first two hidden layers set up as feature detectors for different parts
of the previous layer. All the neurons in the block were set up to have the same weights as those from
the previous layer. This is called weight sharing. Each block would sample a different part of the
previous layer’s image. The first hidden layer had 12 blocks of 8×8 neurons, whereas the second
hidden layer had 12 blocks of 4×4 neurons. The third hidden layer was fully connected and consisted
Applications 294
of 30 neurons. There were 1256 neurons. The network was trained on 7300 examples and tested on
2000 cases with error rates of 1% on training set and 5% on the test set.
Applications 295

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Summary
You explored further the backpropagation algorithm in this chapter, continuing the discussion in Chapter 7.
• A momentum term was added to the training law and was shown to result in much faster
convergence in some cases.
• A noise term was added to inputs to allow training to take place with random noise applied. This
noise was made to decrease with the number of cycles, so that final stage learning could be done in a
noise−free environment.
• The final version of the backpropagation simulator was constructed and used on the example from
Chapter 12. Further application of the simulator will be made in Chapter 14.
• Several applications with the backpropagation algorithm were outlined, showing the wide
applicability of this algorithm.
Summary 296

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Chapter 14
Application to Financial Forecasting
Introduction
In Chapters 7 and 13, the backpropagation simulator was developed. In this chapter, you will use the simulator
to tackle a complex problem in financial forecasting. The application of neural networks to financial
forecasting and modeling has been very popular over the last few years. Financial journals and magazines
frequently mention the use of neural networks, and commercial tools and simulators are quite widespread.
This chapter gives you an overview of typical steps used in creating financial forecasting models. Many of the
steps will be simplified, and so the results will not, unfortunately, be good enough for real life application.
However, this chapter will hopefully serve as an introduction to the field with some added pointers for further
reading and resources for those who want more detailed information.
Who Trades with Neural Networks?

There has been a great amount of interest on Wall Street for neural networks. Bradford Lewis runs two
Fidelity funds in part with the use of neural networks. Also, LBS Capital Management (Peoria, Illinois)
manages part of its portfolio with neural networks. According to Barron’s (February 27, 1995), LBS’s $150
million fund beat the averages by three percentage points a year since 1992. Each weekend, neural networks
are retrained with the latest technical and fundamental data including P/E ratios, earnings results and interest
rates. Another of LBS’s models has done worse than the S&P 500 for the past five years however. In the book
Virtual Trading, Jeffrey Katz states that most of the successful neural network systems are proprietary and not
publicly heard of. Clients who use neural networks usually don’t want anyone else to know what they are
doing, for fear of losing their competitive edge. Firms put in many person−years of engineering design with a
lot of CPU cycles to achieve practical and profitable results. Let’s look at the process:
Developing a Forecasting Model

There are many steps in building a forecasting model, as listed below.
1. Decide on what your target is and develop a neural network (following these steps) for each
target.
2. Determine the time frame that you wish to forecast.
3. Gather information about the problem domain.
4. Gather the needed data and get a feel for each inputs relationship to the target.
5. Process the data to highlight features for the network to discern.
6. Transform the data as appropriate.
7. Scale and bias the data for the network, as needed.
8. Reduce the dimensionality of the input data as much as possible.
Chapter 14 Application to Financial Forecasting 297

9. Design a network architecture (topology, # layers, size of layers, parameters, learning paradigm).
10. Go through the train/test/redesign loop for a network.
11. Eliminate correlated inputs as much as possible, while in step 10.
12. Deploy your network on new data and test it and refine it as necessary.
Once you develop a forecasting model, you then must integrate this into your trading system. A neural
network can be designed to predict direction, or magnitude, or maybe just turning points in a particular market
or something else. Avner Mandelman of Cereus Investments (Los Altos Hills, California) uses a long−range
trained neural network to tell him when the market is making a top or bottom (Barron’s, December 14, 1992).
Now let’s expand on the twelve aspects of model building:
The Target and the Timeframe
What should the output of your neural network forecast? Let’s say you want to predict the stock market. Do
you want to predict the S&P 500? Or, do you want to predict the direction of the S&P 500 perhaps? You
could predict the volatility of the S&P 500 too (maybe if you’re an options player). Further, like Mr.
Mandelman, you could only want to predict tops and bottoms, say, for the Dow Jones Industrial Average. You
need to decide on the market or markets and also on your specific objectives.
Another crucial decision is the timeframe you want to predict forward. It is easier to create neural network
models for longer term predictions than it is for shorter term predictions. You can see a lot of market noise, or
seemingly random, chaotic variations at smaller and smaller timescale resolutions that might explain this.
Another reason is that the macroeconomic forces that fundamentally move market over long periods, move
slowly. The U.S. dollar makes multiyear trends, shaped by economic policy of governments around the world.
For a given error tolerance, a one−year forecast, or one−month forecast will take less effort with a neural
network than a one−day forecast will.
Domain Expertise
So far we’ve talked about the target and the timeframe. Now one other important aspect of model building is
knowledge of the domain. If you want to create an effective predictive model of the weather, then you need to
know or be able to guess about the factors that influence weather. The same holds true for the stock market or
other financial market. In order to create a real tradable Treasury bond trading system, you need to have a
good idea of what really drives the market and works— i.e., talk to a Tbond trader and encapsulate his domain
expertise!
Gather the Data
Once you know the factors that influence the target output, you can gather raw data. If you are predicting the
S&P 500, then you may consider Treasury yields, 3−month T−bill yields, and earnings as some of the factors.
Once you have the data, you can do scatter plots to see if there is some correlation between the input and the
target output. If you are not satisfied with the plot, you may consider a different input in its place.
The Target and the Timeframe 298


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Pre processing the Data for the Network
Surprising as it may sound, you are most likely going to spend about 90% of your time, as a neural network
developer, in massaging and transforming data into meaningful form for training your network. We actually
defined three substeps in this area of preprocessing in our master list:
• Highlight features
• Transform
• Scale and bias
Highlighting Features in the Input Data

You should present the neural network, as much as possible, with an easy way to find patterns in your data.
For time series data, like stock market prices over time, you may consider presenting quantities like rate of
change and acceleration (the first and second derivatives of your input) as examples. Other ways to highlight
data is to magnify certain occurrences. For example, if you consider Central bank intervention as an important
qualifier to foreign exchange rates, then you may include as an input to your network, a value of 1 or 0, to
indicate the presence or lack of presence of Central bank intervention. Now if you further consider the activity
of the U.S. Federal Reserve bank to be important by itself, then you may wish to highlight that, by separating
it out as another 1/0 input. Using 1/0 coding to separate composite effects is called thermometer encoding.
There is a whole body of study of market behavior called Technical Analysis from which you may also wish
to present technical studies on your data. There is a wide assortment of mathematical technical studies that
you perform on your data (see references), such as moving averages to smooth data as an example. There are
also pattern recognition studies you can use, like the “double−top” formation, which purportedly results in a
high probability of significant decline. To be able to recognize such a pattern, you may wish to present a
mathematical function that aids in the identification of the double−top.
You may want to de−emphasize unwanted noise in your input data. If you see a spike in your data, you can
lessen its effect, by passing it through a moving average filter for example. You should be careful about
introducing excessive lag in the resulting data though.
Transform the Data If Appropriate

For time series data, you may consider using a Fourier transform to move to the frequency−phase plane. This
will uncover periodic cyclic information if it exists. The Fourier transform will decompose the input discrete
data series into a series of frequency spikes that measure the relevance of each frequency component. If the
stock market indeed follows the so−called January effect, where prices typically make a run up, then you
would expect a strong yearly component in the frequency spectrum. Mark Jurik suggests sampling data with
intervals that catch different cycle periods, in his paper on neural network data preparation (see references ).
Pre processing the Data for the Network 299

You can use other signal processing techniques such as filtering. Besides the frequency domain, you can also
consider moving to other spaces, such as with using the wavelet transform. You may also analyze the chaotic
component of the data with chaos measures. It’s beyond the scope of this book to discuss these techniques.
(Refer to the Resources section of this chapter for more information.) If you are developing short−term
trading neural network systems, these techniques may play a significant role in your preprocessing effort. All
of these techniques will provide new ways of looking at your data, for possible features to detect in other
domains.
Scale Your Data

Neurons like to see data in a particular input range to be most effective. If you present data like the S&P 500
that varies from 200 to 550 (as the S&P 500 has over the years) will not be useful, since the middle layer of
neurons have a Sigmoid Activation function that squashes large inputs to either 0 or +1. In other words, you
should choose data that fit a range that does not saturate, or overwhelm the network neurons. Choosing inputs
from –1 to 1 or 0 to 1 is a good idea. By the same token, you should normalize the expected values for the
outputs to the 0 to 1 sigmoidal range.
It is important to pay attention to the number of input values in the data set that are close to zero. Since the
weight change law is proportional to the input value, then a close to zero input will mean that that weight will
not participate in learning! To avoid such situations, you can add a constant bias to your data to move the data
closer to 0.5, where the neurons respond very well.
Reduce Dimensionality
You should try to eliminate inputs wherever possible. This will reduce the dimensionality of the problem and
make it easier for your neural network to generalize. Suppose that you have three inputs, x, y and z and one
output, o. Now suppose that you find that all of your inputs are restricted only to one plane. You could
redefine axes such that you have x’ and y’ for the new plane and map your inputs to the new coordinates. This
changes the number of inputs to your problem to 2 instead of 3, without any loss of information. This is
illustrated in Figure 14.1.
Figure 14.1 Reducing dimensionality from three to two dimensions.
Reduce Dimensionality 300


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95

If your overall goal is beyond pattern classification, you need to track your network’s ability to generalize.
Not only should you look at the overall error with the training set that you define, but you should set aside
some training examples as part of a test set (and do not train with them), with which you can see whether or
not the network is able to correctly predict. If the network responds poorly to your test set, you know that you
have overtrained, or you can say the network “memorized” the training patterns. If you look at the arbitrary
curve−fitting analogy in Figure 14.2, you see curves for a generalized fit, labeled G, and an overfit, labeled O.
In the case of the overfit, any data point outside of the training data results in highly erroneous prediction.
Your test data will certainly show you large error in the case of an overfitted model.
Figure 14.2 General (G) versus over fitting (0) of data.
Another way to consider this issue is in terms of Degrees Of Freedom (DOF). For the polynomial:
y= a0 + a1x + a2x2 + anxn...
the DOF equals the number of coefficients a0, a1 ... an, which is N + 1. So for the equation of a line (y=a0 +
a1x), the DOF would be 2. For a parabola, this would be 3 and so on. The objective to not overfit data can be
restated as an objective to obtain the function with the least DOF that fits the data adequately. For neural
network models, the larger the number of trainable weights (which is a function of the number of inputs and
the architecture), the larger the DOF. Be careful with having too many (unimportant) inputs. You may find
terrific results with your training data, but extremely poor results with your test data.
Eliminate Correlated Inputs Where Possible
You have seen that getting to the minimum number of inputs for a given problem is important in terms of
minimizing DOF and simplifying your model. Another way to reduce dimensionality is to look for correlated
inputs and to carefully eliminate redundancy. For example, you may find that the Swiss franc and German
mark are highly correlated over a certain time period of interest. You may wish to eliminate one of these
inputs to reduce dimensionality. You have to be careful in this process though. You may find that a seemingly
redundant piece of information is actually very important. Mark Jurik, of Jurik Consulting, in his paper on
data preprocessing, suggests that one of the best ways to determine if an input is really needed is to construct
neural network models with and without the input and choose the model with the best error on training and
test data. Although very iterative, you can try eliminating as many inputs as possible this way and be assured
that you haven’t eliminated a variable that really made a difference.
Another approach is sensitivity analysis, where you vary one input a little, while holding all others constant
and note the effect on the output. If the effect is small you eliminate that input. This approach is flawed
Eliminate Correlated Inputs Where Possible 301

because in the real world, all the inputs are not constant. Jurik’s approach is more time consuming but will
lead to a better model.
The process of decorrelation, or eliminating correlated inputs, can also utilize a linear algebra technique
called principal component analysis. The result of principal component analysis is a minimum set of variables
that contain the maximum information. For further information on principal component analysis, you should
consult a statistics reference or research two methods of principal component analysis: the Karhunen−Loev
transform and the Hotelling transform.
Design a Network Architecture
Now it’s time to actually design the neural network. For the backpropagation feed−forward neural network we
have designed, this means making the following choices:
1. The number of hidden layers.

2. The size of hidden layers.
3. The learning constant, beta([beta]).
4. The momentum parameter, alpha([alpha]).
5. The form of the squashing function (does not have to be the sigmoid).
6. The starting point, that is, initial weight matrix.
7. The addition of noise.
Some of the parameters listed can be made to vary with the number of cycles executed, similar to the current
implementation of noise. For example, you can start with a learning constant [beta] that is large and reduce
this constant as learning progresses. This allows rapid initial learning in the beginning of the process and may
speed up the overall simulation time.
The Train/Test/Redesign Loop
Much of the process of determining the best parameters for a given application is trial and error. You need to
spend a great deal of time evaluating different options to find the best fit for your problem. You may literally
create hundreds if not thousands of networks either manually or automatically to search for the best solution.
Many commercial neural network programs use genetic algorithms to help to automatically arrive at an
optimum network. A genetic algorithm makes up possible solutions to a problem from a set of starting genes.
Analogous to biological evolution, the algorithm combines genetic solutions with a predefined set of operators
to create new generations of solutions, who survive or perish depending on their ability to solve the problem.
The key benefit of genetic algorithms (GA) is the ability to traverse an enormous search space for a possibly
optimum solution. You would program a GA to search for the number of hidden layers and other network
parameters, and gradually evolve a neural network solution. Some vendors use a GA only to assign a starting
set of weights to the network, instead of randomizing the weights to start you off near a good solution.
Design a Network Architecture 302


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Now let’s review the steps:
1. Split your data. First, divide you data set into three pieces, a training set, a test set and a blind
test set. Use about 80% of your data records for your training set, 10% for your test set and 10% for
your blind test set.
2. Train and test. Next, start with a network topology and train your network on your training set
data. When you have reached a satisfactory minimum error, save your weights and apply your trained
network to the test data and note the error. Now restart the process with the same network topology
for a different set of initial weights and see if you can achieve a better error on training and test sets.
Reasoning: you may have found a local minimum on your first attempt and randomizing the initial
weights will start you off to a different, maybe better solution.
3. Eliminate correlated inputs. You may optionally try at this point to see if you can eliminate
correlated inputs, as mentioned before, by iteratively removing each input and noting the best error
you can achieve on the training and test sets for each of these cases. Choose the case that leads to the
best error and eliminate the input (if any) that achieved it. You can repeat this whole process again to
try to eliminate another input variable.
4. Iteratively train and test. Now you can try other network parameters and repeat the train and
test process to achieve a better result.
5. Deploy your network. You now can use the blind test data set to see how your optimized
network performs. If the error is not satisfactory, then you need to re−enter the design phase or the
train and test phase.
6. Revisit your network design when conditions change. You need to retrain your network when
you have reason to think that you have new information relevant to the problem you are modeling. If
you have a neural network that tries to predict the weekly change in the S&P 500, then you likely will
need to retrain your network at least once a month, if not once a week. If you find that the network no
longer generalizes well with the new information, you need to re−enter the design phase.
If this sounds like a lot of work, it is! Now, let’s try our luck at forecasting by going through a subset of the
steps outlined:
Forecasting the S&P 500

The S&P 500 index is a widely followed stock average, like the Dow Jones Industrial Average (DJIA). It has
a broader representation of the stock market since this average is based on 500 stocks, whereas the DJIA is
based on only 30. The problem to be approached in this chapter is to predict the S&P 500 index, given a
variety of indicators and data for prior weeks.
Choosing the Right Outputs and Objective
Our objective is to forecast the S&P 500 ten weeks from now. Whereas the objective may be to predict the
level of the S&P 500, it is important to simplify the job of the network by asking for a change in the level
rather than for the absolute level of the index. What you want to do is give the network the ability to fit the
Forecasting the S&P 500 303

problem at hand conveniently in the output space of the output layer. Practically speaking, you know that the
output from the network cannot be outside of the 0 to 1 range, since we have used a sigmoid activation
function. You could take the S&P 500 index and scale this absolute price level to this range, for example.
However, you will likely end up with very small numbers that have a small range of variability. The
difference from week to week, on the other hand, has a much smaller overall range, and when these
differences are scaled to the 0 to 1 range, you have much more variability.
The output we choose is the change in the S&P 500 from the current week to 10 weeks from now as a
percentage of the current week’s value.
Choosing the Right Inputs
The inputs to the network need to be weekly changes of indicators that have some relevance to the S&P 500
index. This is a complex forecasting problem, and we can only guess at some of the relationships. This is one
of the inherent strengths of using neural nets for forecasting; if a relationship is weak, the network will learn
to ignore it automatically. Be cognizant that you do want to minimize the DOF as mentioned before though.
In this example, we choose a data set that represents the state of the financial markets and the economy. The
inputs chosen are listed as:
• Previous price action in the S&P 500 index, including the close or final value of the index
• Breadth indicators for the stock market, including the number of advancing issues and declining
issues for the stocks in the New York Stock Exchange (NYSE)
• Other technical indicators, including the number of new highs and new lows achieved in the week
for the NYSE market. This gives some indication about the strength of an uptrend or downtrend.
• Interest rates, including short−term interest rates in the Three−Month Treasury Bill Yield, and
long−term rates in the 30−Year Treasury Bond Yield.
Other possible inputs could have been government statistics like the Consumer Price Index, Housing starts,
and the Unemployment Rate. These were not chosen because long− and short−term interest rates tend to
encompass this data already.
You are encouraged to experiment with other inputs and ideas. All of the data mentioned can be obtained in
the public domain, such as from financial publications (Barron’s, Investor’s Business Daily, Wall Street
Journal) and from ftp sites on the Internet for the Department of Commerce and the Federal Reserve, as well
as from commercial vendors (see the Resource Guide at the end of the chapter). There are new sources
cropping up on the Internet all the time. A sampling of World Wide Web addresses for commercial and
noncommercial sources include:
• FINWeb, http://riskweb.bus.utexas.edu/finweb.html
• Chicago Mercantile Exchange, http://www.cme.com/cme
• SEC Edgar Database, http://town.hall.org/edgar/edgar.html
• CTSNET Business & Finance Center, http://www.cts.com/cts/biz/
• QuoteCom, http://www.quote.com
• Philadelphia Fed, http://compstat.wharton.upenn.edu:8001/~siler/ fedpage.html
• Ohio state Financial Data Finder, http://cob.ohio−state.edu/dept/fin/ osudata.html
Choosing the Right Inputs 304

Choosing the Right Inputs 305


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Choosing a Network Architecture
The input and output layers are fixed by the number of inputs and outputs we are using. In our case, the output
is a single number, the expected change in the S&P 500 index 10 weeks from now. The input layer size will
be dictated by the number of inputs we have after preprocessing. You will see more on this soon. The middle
layers can be either 1 or 2. It is best to choose the smallest number of neurons possible for a given problem to
allow for generalization. If there are too many neurons, you will tend to get memorization of patterns. We will
use one hidden layer. The size of the first hidden layer is generally recommended as between one−half to
three times the size of the input layer. If a second hidden layer is present, you may have between three and ten
times the number of output neurons. The best way to determine optimum size is by trial and error.
NOTE: You should try to make sure that there are enough training examples for your
trainable weights. In other words, your architecture may be dictated by the number of input
training examples, or facts, you have. In an ideal world, you would want to have about 10 or
more facts for each weight. For a 10−10−1 architecture, there are (10X10 + 10X1 = 110
weights), so you should aim for about 1100 facts. The smaller the ratio of facts to weights, the
more likely you will be undertraining your network, which will lead to very poor
generalization capability.
Preprocessing Data
We now begin the preprocessing effort. As mentioned before, this will likely be where you, the neural
network designer, will spend most of your time.
A View of the Raw Data
Let’s look at the raw data for the problem we want to solve. There are a couple of ways we can start
preprocessing the data to reduce the number of inputs and enhance the variability of the data:
• Use Advances/Declines ratio instead of each value separately.

• Use New Highs/New Lows ratio instead of each value separately.
We are left with the following indicators:
1. Three−Month Treasury Bill Yield

2. 30−Year Treasury Bond Yield
3. NYSE Advancing/Declining issues
4. NYSE New Highs/New Lows
5. S&P 500 closing price
Choosing a Network Architecture 306

Raw data for the period from January 4, 1980 to October 28, 1983 is taken as the training period, for a total of
200 weeks of data. The following 50 weeks are kept on reserve for a test period to see if the predictions are
valid outside of the training interval. The last date of this period is October 19, 1984. Let’s look at the raw
data now. (You get data on the disk available with this book that covers the period from January, 1980 to
December, 1992.) In Figures 14.3 through 14.5, you will see a number of these indicators plotted over the
training plus test intervals:
• Figure 14.3 shows you the S&P 500 stock index.
Figure 14.3 The S&P 500 Index for the period of interest.
• Figure 14.4 shows long−term bonds and short−term 3−month T−bill interest rates.
Figure 14.4 Long−term and short−term interest rates.
• Figure 14.5 shows some breadth indicators on the NYSE, the number of advancing stocks/number
of declining stocks, as well as the ratio of new highs to new lows on the NYSE
Figure 14.5 Breadth indicators on the NYSE: Advancing/Declining issues and New Highs/New
Lows.
A sample of a few lines looks like the following data in Table 14.1. Note that the order of parameters is the
same as listed above.
Table 14.1 Raw Data

Date3Mo TBills30YrTBondsNYSE−Adv/DecNYSE−NewH/NewLSP−Close
1/4/8012.119.644.2094592.764706106.52
1/11/8011.949.731.64957321.28571109.92
1/18/8011.99.80.8813354.210526111.07
1/25/8012.199.930.7932693.606061113.61
2/1/8012.0410.21.162932.088235115.12
2/8/8012.0910.481.3384152.936508117.95
2/15/8012.3110.960.3380530.134615115.41
2/22/8013.1611.250.323810.109091115.04
2/29/8013.712.141.6768950.179245113.66
3/7/8015.1412.10.2825910106.9
3/14/8015.3812.010.6902860.011628105.43
3/21/8015.0511.730.4862670.027933102.31
3/28/8016.5311.675.2471910.011628100.68
4/3/8015.0412.060.9835620.117647102.15
4/11/8014.4211.811.5658540.310345103.79
4/18/8013.8211.231.1132870.146341100.55
4/25/8012.7310.590.8498070.473684105.16
Choosing a Network Architecture 307

5/2/8010.7910.421.1474651.857143105.58
5/9/809.7310.150.5130520.473684104.72
5/16/808.69.71.3424446.75107.35
5/23/808.959.873.11082526110.62
Highlight Features in the Data
For each of the five inputs, we want use a function to highlight rate of change types of features. We will use
the following function (as originally proposed by Jurik) for this purpose.
ROC(n) = (input(t) − BA(t − n)) / (input(t)+ BA(t − n))
where: input(t) is the input’s current value and BA(t − n) is a five unit block average of adjacent values
centered around the value n periods ago.
Now we need to decide how many of these features we need. Since we are making a prediction 10 weeks into
the future, we will take data as far back as 10 weeks also. This will be ROC(10). We will also use one other
rate of change, ROC(3). We have now added 5*2 = 10 inputs to our network, for a total of 15. All of the
preprocessing can be done with a spreadsheet.
Highlight Features in the Data 308


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Here’s what we get (Table 14.2) after doing the block averages. Example : BA3MoBills for 1/18/80 =
(3MoBills(1/4/80) + 3MoBills(1/11/80) + 3MoBills(1/18/80) + 3MoBills(1/25/80) + 3MoBills(2/1/80))/5.
Table 14.2 Data after Doing Block Averages

Date3MoBillsLngBondsNYSE−
1/4/8012.119.644.209459
1/11/8011.949.731.649573
1/18/8011.99.80.881335
1/25/8012.199.930.793269
2/1/8012.0410.21.16293
2/8/8012.0910.481.338415
2/15/8012.3110.960.338053
2/22/8013.1611.250.32381
2/29/8013.712.141.676895
3/7/8015.1412.10.282591
3/14/8015.3812.010.690286
3/21/8015.0511.730.486267
3/28/8016.5311.675.247191
4/3/8015.0412.060.983562
4/11/8014.4211.811.565854
4/18/8013.8211.231.113287
4/25/8012.7310.590.849807
5/2/8010.7910.421.147465
5/9/809.7310.150.513052
5/16/808.69.71.342444
5/23/808.959.873.110825
NYSE−Adv/DecSP−Close NewH/NewLBA3MoBBALngBnd
2.764706106.52
21.28571109.92
4.210526111.0712.0369.86
3.606061113.6112.03210.028
2.088235115.1212.10610.274
2.936508117.9512.35810.564
0.134615115.4112.6611.006
0.109091115.0413.2811.386
0.179245113.6613.93811.692
0106.914.48611.846
0.011628105.4315.1611.93
0.027933102.3115.42811.914
0.011628100.6815.28411.856
0.117647102.1514.97211.7
0.310345103.7914.50811.472

0.146341100.5513.3611.222
0.473684105.1612.29810.84
1.857143105.5811.13410.418
0.473684104.7210.1610.146
6.75107.357.6148.028
26110.625.4565.944
BAA/DBAH/LBAClose
1.7393136.791048111.248
1.1651046.825408113.534
0.90282.595189114.632
0.7912951.774902115.426
0.9680211.089539115.436
0.7919530.671892113.792
0.6623270.086916111.288
0.691970.065579108.668
1.6766460.046087105.796
1.5379790.033767103.494
1.7946320.095836102.872
1.8792320.122779101.896
1.951940.211929102.466
1.1319950.581032103.446
1.0378930.652239103.96
0.9932111.94017104.672
1.3927197.110902106.686
1.2227577.01616585.654
0.9932646.64473764.538


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Now let’s look at the rest of this table, which is made up of the new 10 values of ROC indicators (Table 14.3).
Table 14.3 Added Rate of Change (ROC) Indicators

DateROC3_3MoROC3_BondROC10_ADROC3_H/LROC3_SPC
1/4/80
1/11/80
1/18/80
1/25/80
2/1/80
2/8/800.0022380.030482−0.13026−0.396250.029241
2/15/800.0114210.044406−0.55021−0.961320.008194
2/22/800.0417160.045345−0.47202−0.919320.001776
2/29/800.05150.0694150.358805−0.81655−0.00771
3/7/800.0892090.047347−0.54808−1−0.03839
3/14/800.0732730.026671−0.06859−0.96598−0.03814
3/21/800.0383610.001622−0.15328−0.51357−0.04203
3/28/800.065901−0.007480.766981−0.69879−0.03816
4/3/80−0.003970.005419−0.260540.437052−0.01753
4/11/80−0.03377−0.004380.0089810.437052−0.01753
4/18/80−0.0503−0.02712−0.234310.8037430.001428
4/25/80−0.08093−0.0498−0.377210.588310.015764
5/2/80−0.14697−0.04805−0.259560.7951460.014968
5/9/80−0.15721−0.05016−0.37625−0.101780.00612
5/16/80−0.17695−0.05550.1279440.8237720.016043
5/23/80−0.10874−0.027010.5159830.861120.027628
ROC10_3MoROC10_BndROC10_ADROC10_HLROC10_SP
0.157320.0840690.502093−0.99658−0.04987
0.1111110.091996−0.08449−0.96611−0.05278
0.0872350.0695530.268589−0.78638−0.04964
0.0558480.0305590.169062−0.84766−0.06888
0.002757−0.01926−0.06503−0.39396−0.04658
−0.10345−0.04430.1833090.468658−0.03743
−0.17779−0.0706−0.1270.689919−0.03041

−0.25496−0.09960.3197350.980756−0.0061
−0.25757−0.09450.2995690.9964610.02229
NOTE: Note that you don’t get completed rows until 3/28/90, since we have a ROC
indicator dependent on a Block Average value 10 weeks before it. The first block average
value is generated 1/1/80, two weeks after the start of the data set. All of this indicates that
you will need to discard the first 12 values in the dataset to get complete rows, also called
complete facts.
Normalizing the Range
We now have values in the original five data columns that have a very large range. We would like to reduce
the range by some method. We use the following function:
new value = (old value − Mean)/ (Maximum Range)
This relates the distance from the mean for a value in a column as a fraction of the Maximum range for that
column. You should note the value of the Maximum range and Mean, so that you can un−normalize the data
when you get a result.
The Target
We’ve taken care of all our inputs, which number 15. The final piece of information is the target. The
objective as stated at the beginning of this exercise is to predict the percentage change 10 weeks into the
future. We need to time shift the S&P 500 close 10 weeks back, and then calculate the value as a percentage
change as follows:
Result = 100 X ((S&P 10 weeks ahead) − (S&P this week))/(S&P this week).
This gives us a value that varies between −14.8 to and + 33.7. This is not in the form we need yet. As you
recall, the output comes from a sigmoid function that is restricted to 0 to +1. We will first add 14.8 to all
values and then scale them by a factor of 0.02. This will result in a scaled target that varies from 0 to 1.
scaled target = (result + 14.8) X 0.02
The final data file with the scaled target shown along with the scaled original six columns of data is shown in
Table 14.4.
Table 14.4 Normalized Ranges for Original Columns and Scaled Target
DateS_3MOBillS_LngBndS_A/D
3/28/800.534853−0.016160.765273
4/3/800.3913080.055271−0.06356
4/11/800.3315780.0094830.049635
4/18/800.273774−0.09674−0.03834
4/25/800.168765−0.21396−0.08956
5/2/80−0.01813−0.2451−0.0317
5/9/80−0.12025−0.29455−0.15503
5/16/80−0.22912−0.376960.006205
Normalizing the Range 312

5/23/80−0.1954−0.345830.349971
S_H/LS_SPCResultScaled Target
−0.07089−0.5132812.435440.544709
−0.07046−0.4923612.883020.55366
−0.06969−0.469019.894980.4939
−0.07035−0.5151315.365490.60331
−0.06903−0.4495111.715480.53031
−0.06345−0.4435311.612050.528241
−0.06903−0.4557716.539340.626787
−0.04372−0.4183312.510480.54621
0.033901−0.371799.5733140.487466
Normalizing the Range 313


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Storing Data in Different Files
You need to place the first 200 lines in a training.dat file (provided for you in the accompanying diskette) and
the subsequent 40 lines of data in the another test.dat file for use in testing. You will read more about this
shortly. There is also more data than this provided on this diskette in raw form for you to do further
experiments.
Training and Testing

With the training data available, we set up a simulation. The number of inputs are 15, and the number of
outputs is 1. A total of three layers are used with a middle layer of size 5. This number should be made as
small as possible with acceptable results. The optimum sizes and number of layers can only be found by much
trial and error. After each run, you can look at the error from the training set and from the test set.
Using the Simulator to Calculate Error
You obtain the error for the test set by running the simulator in Training mode (you need to temporarily copy
the test data with expected outputs to the training.dat file) for one cycle with weights loaded from the weights
file. Since this is the last and only cycle, weights do not get modified, and you can get a reading of the average
error. Refer to Chapter 13 for more information on the simulator’s Test and Training modes. This approach
has been taken with five runs of the simulator for 500 cycles each. Table 14.5 summarizes the results along
with the parameters used. The error gets better and better with each run up to run # 4. For run #5, the training
set error decreases, but the test set error increases, indicating the onset of memorization. Run # 4 is used for
the final network results, showing test set RMS error of 13.9% and training set error of 6.9%.
Table 14.5 Results of Training the Backpropagation Simulator for Predicting the Percentage Change in the
S&P 500 Index
Run#ToleranceBetaAlphaNFmax cyclescycles runtraining set errortest set error
10.0010.50.0010.00055005000.1509380.25429
20.0010.40.0010.00055005000.1149480.185828
30.0010.3005005000.09364220.148541
40.0010.2005005000.0689760.139230
50.0010.1005005000.06214120.143430
NOTE: If you find the test set error does not decrease much, whereas the training set error
continues to make substantial progress, then this means that memorization is starting to set in
(run#5 in example). It is important to monitor the test set(s) that you are using while you are
training to make sure that good, generalized learning is occurring versus memorizing or
overfitting the data. In the case shown, the test set error continued to improve until run#5,
where the test set error degraded. You need to revisit the 12−step process to forecasting
model design to make any further improvements beyond what was achieved.
Storing Data in Different Files 314

To see the exact correlation, you can copy any period you’d like, with the expected value output fields
deleted, to the test.dat file. Then you run the simulator in Test mode and get the output value from the
simulator for each input vector. You can then compare this with the expected value in your training set or test
set.
Now that you’re done, you need to un−normalize the data back to get the answer in terms of the change in the
S&P 500 index. What you’ve accomplished is a way in which you can get data from a financial newspaper,
like Barron’s or Investor’s Business Daily, and feed the current week’s data into your trained neural network
to get a prediction of what the S&P 500 index is likely to do ten weeks from now.
Here are the steps to un−normalize:
1. Take the predicted scaled target value and calculate, the result value as Result = (Scaled
target/0.02) − 14.8
2. Take the result from step 1 (which is the percentage change 10 weeks from now) and calculate
the projected value, Projected S&P 10 weeks from now = (This week’s S&P value)(1+ Result/100)
Only the Beginning
This is only a very brief illustration (not meant for trading !) of what you can do with neural networks in
financial forecasting. You need to further analyze the data, provide more predictive indicators, and
optimize/redesign your neural network architecture to get better generalization and lower error. You need to
present many, many more test cases representing different market conditions to have a robust predictor that
can be traded with. A graph of the expected and predicted output for the test set and the training set is shown
in Figure 14.6. Here, the normalized values are used for the output. Note that the error is about 13.9% on
average over the test set and 6.9% over the training set. You can see that the test set did well in the beginning,
but showed great divergence in the last few weeks.
Figure 14.6 Comparison of predicted versus actual for the training and test data sets.
The preprocessing steps shown in this chapter should serve as one example of the kinds of steps you can use.
There are a vast variety of analysis and statistical methods that can be used in preprocessing. For applying
fuzzy data, you can use a program like the fuzzifier program that was developed in Chapter 3 to preprocess
some of the data.
What’s Next?
There are many other experiments you can do from here on. The example chosen was in the field of financial
forecasting. But you could certainly try the simulator on other problems like sales forecasting or perhaps even
weather forecasting. The key to all of the applications though, is how you present and enhance data, and
working through parameter selection by trial and error. Before concluding this chapter, we will cover more
topics in preprocessing and present some case studies in financial forecasting. You should consider the
suggestions made in the 12−step approach to forecasting model design and research some of the resources
listed at the end of this chapter if you have more interest in this area.
Only the Beginning 315

Only the Beginning 316


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Technical Analysis and Neural Network Preprocessing

We cannot overstate the importance of preprocessing in developing a forecasting model. There is a large body
of information related to the study of financial market behavior called Technical Analysis. You can use the
mathematical studies defined by Technical Analysis to preprocess your input data to reveal predictive
features. We will present a sampling of Technical Analysis studies that can be used, with formulae and
graphs.
Moving Averages
Moving averages are used very widely to capture the underlying trend of a price move. Moving averages are
simple filters that average data over a moving window. Popular moving averages include 5−, 10−, and 20
period moving averages. The formula is shown below for a simple moving average, SMA:
SMAt = ( Pt + Pt−1 + ... Pt−n)/ n

where n = the number of time periods back
P−n= price at n time periods back
An exponential moving average is a weighted moving average that places more weight on the most recent
data. The formula for this indicator, EMA is as follows:
EMAt = (1 − a)Pt + a ( EMAt−1)

where a = smoothing constant (typical 0.10)
Pt= price at time t
Momentum and Rate of Change
Momentum is really velocity, or rate of price change with time. The formula for this is
Mt = ( Pt − Pt−a )
where a = lookback parameter
for a 5−day momentum value, a = 5
The Rate of Change indicator is actually a ratio. It is the current price divided by the price some interval, a,
ago divided by a constant. Specifically,
ROC = Pt / Pt−a x 1000
Relative Strength Index
The Relative Strength Index (RSI) is the strength of up closes versus down closes over a certain time interval.
It is calculated over a time interval T as :
RSI = 100 − [ 100 / (1 + RS )]
Technical Analysis and Neural Network Preprocessing 317

where RS = average of x days’ up closes/ average of x days’ down

closes
A typical time interval, T, is 14 days. The assumption with the use of RSI is that higher values up closes
relative to down closes indicates a strong market, and the opposite indicates weak markets.
Percentage R
This indicator measures where in a recent range of prices today’s price falls. The indicator assumes that prices
regress to their mean. A low %R indicates that prices are hitting the ceiling of a range, while a high %R
indicates that prices are at their low in a range. The formula is:
%R = 100 x (HighX − P)/(HighX − LowX)

where HighX is the highest price over the price interval of interest
LowX is the lowest price over the price interval of interest
P is the current price
Herrick Payoff Index
This indicator makes use of other market data that is available besides price information. It uses the volume of
the security, which, for a stock, is the number of shares traded for a stock during a particular interval. It also
uses the open interest, which is the value of the total number of open trades at a particular time. For a
commodity future, this is the number of open short and long positions. This study attempts to measure the
flow of money in and out of a market. The formula for this is as follows (note that a tick is the smallest
permissible move in a given market) :
Let MP = mean price over a particular interval

OI = the larger of yesterday’s or today’s open interest
then
K =[ (MPt − MPt−1 ) x dollar value of 1 tick move x volume ]

x [1 +/− 2/OI]
HPIt = HPIt−1 + [ 0.1 x (K − HPIt−1 )] / 100,000
MACD
The MACD (moving average convergence divergence) indicator is the difference between two moving
averages, and it tells you when short−term overbought or oversold conditions exist in the market. The formula
is as follows:
Let OSC = EMA1 − EMA2,

where EMA1 is for one smoothing constant and time period, for example
0.15 and 12 weeks
EMA2 is for another smoothing constant and time period,
for example
0.075 and 26 weeks
then
MACDt = MACDt−1 + K x ( OSCt − MACDt−1 )

where K is a smoothing constant, for example, 0.2
Percentage R 318
The final formula effectively does another exponential smoothing on the difference of the two moving
averages, for example, over a 9−week period.
“Stochastics”
This indicator has absolutely nothing to do with stochastic processes. The reason for the name is a mystery,
but the indicator is composed of two parts: %K and %D, which is a moving average of %K. The crossover of
these lines indicates overbought and oversold areas. The formulas follow:
Raw %K = 100 x (P − LowX )/(HighX − LowX)

%Kt = [( %Kt−1 x 2 ) + Raw %Kt ] /3
%Dt = [( %Dt−1 x 2 ) + %Kt ] /3
“Stochastics” 319

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
On−Balance Volume
The on−balance volume (OBV) indicator was created to try to uncover accumulation and distribution patterns
of large player in the stock market. This is a cumulative sum of volume data, specified as follows:
If today’s close is greater than yesterday’s close
OBVt = OBVt−1 + 1
If today’s close is less than yesterday’s close
OBVt = OBVt−1 − 1
The absolute value of the index is not important; attention is given only to the direction and trend.
Accumulation−Distribution
This indicator does for price what OBV does for volume.
If today’s close is greater than yesterday’s close:
ADt = ADt−1 + (Closet − Lowt)
If today’s close is less than yesterday’s close
ADt = ADt−1 + (Hight − Closet)
Now let’s examine how these indicators look. Figure 14.7 shows a bar chart, which is a chart of price data
versus time, along with the following indicators:
• Ten−unit moving average

• Ten−unit exponential moving average
• Momentum
• MACD
• Percent R
Figure 14.7 Five minute bar chart of the S&P 500 Sept 95 Futures contract with several technical indicators
displayed.
On−Balance Volume 320

The time period shown is 5 minute bars for the S&P 500 September 1995 Futures contract. The top of each
bar indicates the highest value (“high”) for that time interval, the bottom indicates the lowest value(“low”),
and the horizontal lines on the bar indicate the initial (“open”) and final (“close”) values for the time interval.
Figure 14.8 shows another bar chart for Intel Corporation stock for the period from December 1994 to July
1995, with each bar representing a day of activity. The following indicators are displayed also.
• Rate of Change
• Relative Strength
• Stochastics
• Accumulation−Distribution
Figure 14.8 Daily bar chart of Intel Corporation with several technical indicators displayed.
You have seen a few of the hundreds of technical indicators that have been invented to date. New indicators
are being created rapidly as the field of Technical Analysis gains popularity and following. There are also
pattern recognition studies, such as formations that resemble flags or pennants as well as more exotic types of
studies, like Elliot wave counts. You can refer to books on Technical Analysis (e.g., Murphy) for more
information about these and other studies.
Neural preprocessing with Technical Analysis tools as well as with traditional engineering analysis tools such
as Fourier series, Wavelets, and Fractals can be very useful in finding predictive patterns for forecasting.
What Others Have Reported

In this final section of the chapter, we outline some case studies documented in periodicals and books, to give
you an idea of the successes or failures to date with neural networks in financial forecasting. Keep in mind
that the very best (= most profitable) results are usually never reported (so as not to lose a competitive edge) !
Also, remember that the market inefficiencies exploited yesterday may no longer be the same to exploit today.
Can a Three−Year−Old Trade Commodities?
Well, Hillary Clinton can certainly trade commodities, but a three−year−old, too? In his paper, “Commodity
Trading with a Three Year Old,” J. E. Collard describes a neural network with the supposed intelligence of a
three−year−old. The application used a feedforward backpropagation network with a 37−30−1 architecture.
The network was trained to buy (“go long”) or sell (“go short”) in the live cattle commodity futures market.
The training set consisted of 789 facts for trading days in 1988, 1989, 1990, and 1991. Each input vector
consisted of 18 fundamental indicators and six market technical variables (Open, High, Low, Close, Open
Interest, Volume). The network could be trained for the correct output on all but 11 of the 789 facts.
The fully trained network was used on 178 subsequent trading days in 1991. The cumulative profit increased
from $0 to $1547.50 over this period by trading one live cattle contract. The largest loss in a trade was
$601.74 and the largest gain in a trade was $648.30.
What Others Have Reported 321

Forecasting Treasury Bill and Treasury Note Yields
Milam Aiken designed a feedforward backpropagation network that predicted Treasury Bill Rates and
compared the forecast he obtained with forecasts made by top U.S. economists. The results showed the neural
network, given the same data, made better predictions (.18 versus .71 absolute error). Aiken used 250
economic data series to see correlation to T−Bills and used only the series that showed leading correlation:
Dept. of Commerce Index of Leading Economic Indicators, the Center for International Business Cycle
Research (CIBCR) Short Leading Composite Index, and the CIBCR Long Leading Composite Index. Prior
data for these three indicators for the past four years (total 12 inputs) was used to predict the average annual
T−Bill rate (one output) for the current year.
Guido Deboeck and Masud Cader designed profitable trading systems for two−year and 10−year treasury
securities. They used feedforward neural networks with a learning algorithm called
extended−delta−bar−delta (EDBD), which is a variant of backpropagation. Training samples composed of
100 facts were selected from 1120 trading days spanning from July 1 1989 to June 30, 1992. The test period
consisted of more than 150 trading days from July 1, 1992 to December 30, 1992. Performance on the test set
was monitored every N thousand training cycles, and the training procedure was stopped when performance
degraded on the test set. (This is the same procedure we used when developing a model for the S&P 500.)
A criterion used to judge model performance was the ratio of the average profit divided by the maximum
drawdown, which is the largest unrealized loss seen during the trading period. A portfolio of separate
designed trading systems for two−year and 10−year securities gave the following performance: Over a period
of 4.5 years, the portfolio had 133 total trades with 65% profitable trades and the maximum drawdown of 64
basis points, or thousands of units for bond yields. The total gain was 677 basis points over that period with a
maximum gain in one trade of 52 basis points and maximum loss in one trade of 47 basis points.
Forecasting Treasury Bill and Treasury Note Yields 322


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
The stability and robustness of this system was checked by using over 1000 moving time windows of
3−month, 6−month, and 12−month duration over the 4.5−year interval and noting the standard deviations in
profits and maximum drawdown. The maximum drawdown varied from 30 to 48 basis points.
Neural Nets versus Box−Jenkins Time−Series Forecasting
Ramesh Sharda and Rajendra Patil used a standard 12−12−1 feedforward backpropagation network and
compared the results with Box−Jenkins methodology for time−series forecasting. Box−Jenkins forecasting is
traditional time−series forecasting technique. The authors used 75 different time series for evaluation. The
results showed that neural networks achieved better MAPE (mean absolute percentage error) with a mean
over all 75 time series MAPEs of 14.67 versus 15.94 for the Box−Jenkins approach.
Neural Nets versus Regression Analysis
Leorey Marquez et al. compared neural network modeling with standard regression analysis. The authors used
a feedforward backpropagation network with a structure of 1−6−1. They used three functional forms found in
regression analysis:
1. Y = B0 + B1 X + e
2. Y = B0 + B1 log(X) + e
3. Y = B0 + B1/X + e
For each of these forms, 100 pairs of (x,y) data were generated for this “true” model.
Now the neural network was trained on these 100 pairs of data. An additional 100 data points were generated
by the network to test the forecasting ability of the network. The results showed that the neural network
achieved a MAPE within 0.6% of the true model, which is a very good result. The neural network model
approximated the linear model best. An experiment was also done with intentional mis−specification of some
data points. The neural network model did well in these cases also, but comparatively worse for the reciprocal
model case.
Hierarchical Neural Network
Mendelsohn developed a multilevel neural network as shown in Figure 14.9. Here five neural networks are
arranged such that four network outputs feed that final network. The four networks are trained to produce the
High, Low, short−term trend, and medium−term trend for a particular financial instrument. The final network
takes these four outputs as input and produces a turning point indicator.
Figure 14.9 Hierarchical neural network system to predict turning points.
Neural Nets versus Box−Jenkins Time−Series Forecasting 323

Each network was trained and tested with 1200 fact days spanning 1988 to 1992 (33% used for testing).
Preprocessing was accomplished by using differences of the inputs and with some technical analysis studies:
• Moving averages
• Exponential moving averages
• Stochastic indicators
For the network that produces a predicted High value, the average error ranged between 7.04% and 7.65% for
various financial markets over the test period, including Treasury Bonds, Eurodollar, Japanese Yen, and S&P
500 futures contracts.
The Walk−Forward Methodology of Market Prediction
A methodology that is sometimes used in neural network design is walk−forward training and testing. This
means that you choose an interval of time (e.g., six months) over which you train a neural network and test the
network over a subsequent interval. You then move the training window and testing window forward one
month, for example, and repeat the exercise. You do this for the time period of interest to see your forecasting
results. The advantage of this approach is that you maximize the network’s ability to model the recent past in
making a prediction. The disadvantage is that the network forget characteristics of the market that happened
prior to the training window.
Takashi Kimoto et al. used the walk forward methodology in designing a trading system for Fujitsu and Nikko
Securities. They also, like Mendelsohn, use a hierarchical neural network composed of individual feedforward
neural networks. Prediction of the TOPIX, which is the Japanese equivalent of the Dow Jones Industrial
Average, was performed for 33 months from January 1987 to September 1980. Four networks were used in
the first level of the hierarchy trained on price data and economic data. The results were fed to a final network
that generated buy and sell signals. The performance of the trading system achieved a result that is 20% better
than a buy and hold strategy for the TOPIX.
Dual Confirmation Trading System
Jeremy Konstenius, discusses a trading system for the S&P 400 index with a holographic neural network,
which is unlike the feedforward backpropagation neural network. The holographic network uses complex
numbers for data input and output from neurons, which are mathematically more complex than feedforward
network neurons. The author uses two trained networks to forecast the next day’s direction based on data for
the past 10 days. Each network uses input data that is detrended, by subtracting a moving average from the
data. Network 1 uses detrended closing values. Network 2 uses detrended High values. If both networks agree,
or confirm each other, then a trade is made. There is no trade otherwise.
Network 1 showed an accuracy of 61.9% for the five−month test period (the training period spanned two
years prior to the test period), while Network 2 also showed an accuracy of 61.9%. Using the two networks
together, Konstenius achieved an accuracy of 65.82%.
A Turning Point Predictor
This neural network approach is discussed by Michitaka Kosaka et al. (1991).
They discuss applying the feedforward backpropagation network to develop buy/sell signals for securities.
You would gather time−series data on stock prices, and want to find trends in the data so that changes in the
direction of the trend provide you the turning points, which you interpret as signals to buy or sell.
The Walk−Forward Methodology of Market Prediction 324

You will need to list these factors that you think have any influence on the price of a security you are
studying. You need to also determine how you measure these factors. You then formulate a nonlinear function
combining the factors on your list and the past however many prices of your security (your time series data).
The function has the form, as Michitaka Kosaka, et al. (1991) put it,
p(t + h) = F(x(t), x(t −1), ... , f1, f2, ... )

where
f1, f2, represent factors on your list,
x(t) is the price of your stock at time t,
p(t + h) is the turning point of security price at time t + h, and
p(t + h) = −1 for a turn from downward to upward,
p(t + h) = +1 for a turn from upward to downward,
p(t + h) = 0 for no change and therefore no turn
Here you vary h through the values 1, 2, etc. as you move into the future one day (time period) at a time. Note
that the detailed form of the function F is not given. This is for you to set up as you see fit.
The Walk−Forward Methodology of Market Prediction 325


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
You can set up a similar function for x(t + h), the stock price at time t + h, and have a separate network
computing it using the backpropagation paradigm. You will then be generating future prices of the stock and
the future buy/sell signals hand in hand, but parallel.
Michitaka Kosaka, et al. (1991) report that they used time−series data over five years to identify the network
model, and time−series data over one year to evaluate the model’s forecasting performance, with a success
rate of 65% for turning points.
The S&P 500 and Sunspot Predictions
Michael Azoff in his book on time−series forecasting with neural networks (see references) creates neural
network systems for predicting the S&P 500 index as well as for predicting chaotic time series, such as
sunspot occurrences. Azoff uses feedforward backpropagation networks, with a training algorithm called
adaptive steepest descent, a variation of the standard algorithm. For the sunspot time series, and an
architecture of 6−5−1, and a ratio of training vectors to trainable weights of 5.1, he achieves training set error
of 12.9% and test set error of 21.4%. This series was composed of yearly sunspot numbers for the years 1706
to 1914. Six years of consecutive annual data were input to the network.
One network Azoff used to forecast the S&P 500 index was a 17−7−1 network. The training vectors to
trainable weights ratio was 6.1. The achieved training set error was 3.29%, and on the test set error was
4.67%. Inputs to this network included price data, a volatility indicator, which is a function of the range of
price movement, and a random walk indicator, a technical analysis study.
A Critique of Neural Network Time−Series Forecasting for Trading
Michael de la Maza and Deniz Yuret, managers for the Redfire Capital Management Group, suggest that
risk−adjusted return, and not mean−squared error should be the metric to optimize in a neural network
application for trading. They also point out that with neural networks, like with statistical methods such as
linear regression, data facts that seem unexplainable can’t be ignored even if you want them to be. There is no
equivalent for a “don’t care,” condition for the output of a neural network. This type of condition may be an
important option for trading environments that have no “discoverable regularity” as the authors put it, and
therefore are really not tradable. Some solutions to the two problems posed are given as follows:
• Use an algorithm other than backpropagation, which allows for maximization of risk−adjusted
return, such as simulated annealing or genetic algorithms.
• Transform the data input to the network so that minimizing mean−squared error becomes
equivalent to maximizing risk−adjusted return.
• Use a hierarchy (see hierarchical neural network earlier in this section) of neural networks, with
each network responsible for detecting features or regularities from one component of the data.
The S&P 500 and Sunspot Predictions 326

Resource Guide for Neural Networks and Fuzzy Logic in Finance

Here is a sampling of resources compiled from trade literature:
NOTE: We do not take responsibility for any errors or omissions.
Magazines
Technical Analysis of Stocks and Commodities

Technical Analysis, Inc., 3517 S.W. Alaska St., Seattle, WA 98146.
Futures
Futures Magazine, 219 Parkade, Cedar Falls, IA 50613.
AI in Finance
Miller Freeman Inc, 600 Harrison St., San Francisco, CA 94107
NeuroVest Journal
P.O. Box 764, Haymarket, VA 22069
IEEE Transactions on Neural Networks
IEEE Service Center, 445 Hoes Lane, P.O. Box 1331, Piscataway, NJ 08855
Particularly worthwhile is an excellent series of articles by consultant Murray Ruggiero Jr., in Futures
magazine on neural network design and trading system design in issues spanning July ‘94 through June ‘95.
Books
Azoff, Michael, Neural Network Time Series Forecasting of Financial Markets, John Wiley and Sons,
New York, 1994.
Lederman, Jess, Virtual Trading, Probus Publishing, 1995.
Trippi, Robert, Neural Networks in Finance and Investing, Probus Publishing, 1993.
Book Vendors
Traders Press, Inc. (800) 927−8222

P.O. Box 6206, Greenville, SC 29606
Traders’ Library (800) 272−2855

9051 Red Branch Rd., Suite M, Columbia, MD 21045
Resource Guide for Neural Networks and Fuzzy Logic in Finance 327

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Consultants
Mark Jurik
Jurik Research
P.O. Box 2379, Aptos, CA 95001
Hy Rao
Via Software Inc, v: (609) 275−4786, fax: (609) 799−7863
BEI Suite 480, 660 Plainsboro Rd., Plainsboro, NJ 08536
ViaSW@aol.com
Mendelsohn Enterprises Inc.

25941 Apple Blossom Lane
Wesley Chapel, FL 33544
Murray Ruggiero Jr.

Ruggiero Associates,
East Haven, CT
The Schwartz Associates (800) 965−4561

801 West El Camino Real, Suite 150, Mountain View, CA 94040
Historical Financial Data Vendors
CSI (800) CSI−4727

200 W. Palmetto Park Rd., Boca Raton, FL 33432
Free Financial Network, New York (212) 838−6324
Genesis Financial Data Services (800) 808−DATA

411 Woodmen, Colorado Springs, CO 80991
Pinnacle Data Corp. (800) 724−4903

460 Trailwood Ct., Webster, NY 14580
Stock Data Corp. (410) 280−5533

905 Bywater Rd., Annapolis, MD 21401
Technical Tools Inc. (800) 231−8005

334 State St., Suite 201, Los Altos, CA 94022
Tick Data Inc. (800) 822−8425
Consultants 328
720 Kipling St., Suite 115, Lakewood, CO 80215
Worden Bros., Inc. (800) 776−4940

4905 Pine Cone Dr., Suite 12, Durham, NC 27707
Preprocessing Tools for Neural Network Development
NeuralApp Preprocessor for Windows

Via Software Inc., v: (609) 275−4786 fax: (609) 799−7863
ViaSW@aol.com
Stock Prophet
Future Wave Software (310) 540−5373
1330 S. Gertruda Ave., Redondo Beach, CA 90277
Wavesamp & Data Decorrelator & Reducer

TSA (800) 965−4561
801 W. El Camino Real, #150, Mountain. View, CA 94040
Genetic Algorithms Tool Vendors
C Darwin
ITANIS International Inc.
1737 Holly Lane, Pittsburgh, PA 15216
EOS
Man Machine Interfaces Inc. (516) 249−4700
555 Broad Hollow Rd., Melville, NY 11747
Evolver
Axcelis, Inc. (206) 632−0885
4668 Eastern Ave. N., Seattle, WA 98103
Fuzzy Logic Tool Vendors
CubiCalc
HyperLogic Corp. (619) 746−2765
1855 East Valley Pkwy., Suite 210, Escondido, CA 92027
TILSHELL
Togai InfraLogic Inc.
5 Vanderbilt, Irvine, CA 92718
Neural Network Development Tool Vendors
Braincel
Promised Land Technologies (203) 562−7335
195 Church St., 8th Floor, New Haven, CT 06510
BrainMaker
Preprocessing Tools for Neural Network Development 329

California Scientific Software (916) 478−9040

10024 Newtown Rd., Nevada City, CA 95959
ForecastAgent for Windows, ForecastAgent for Windows 95

Via Software Inc., v: (609) 275−4786 fax: (609) 799−7863
ViaSW@aol.com
InvestN 32
RaceCom, Inc. (800) 638−8088
555 West Granada Blvd., Suite E−10, Ormond Beach, FL 32714
NetCaster, DataCaster
Maui Analysis & Synthesis Technologies (808) 875−2516
590 Lipoa Pkwy., Suite 226, Kihei, HI 96753
NeuroForecaster
NIBS Pte. Ltd. (65) 344−2357
62 Fowlie Rd., Republic of Singapore 1542
NeuroShell
Ward Systems Group (301)662−7950
Executive Park West, 5 Hillscrest Dr., Frederick, MD 21702
NeuralWorks Predict
NeuralWare Inc. (412) 787−8222
202 Park West Dr., Pittsburgh, PA 15276
N−Train
Scientific Consultant Services (516) 696−3333
20 Stagecoach Rd., Selden, NY 11784
Summary
This chapter presented a neural network application in financial forecasting. As an example of the steps
needed to develop a neural network forecasting model, the change in the Standard & Poor’s 500 stock index
was predicted 10 weeks out based on weekly data for five indicators. Some examples of preprocessing of data
for the network were shown as well as issues in training.
At the end of the training period, it was seen that memorization was taking place, since the error in the test
data degraded, whereas the error in the training set improved. It is important to monitor the error in the test
data (without weight changes) while you are training to ensure that generalization ability is maintained. The
final network resulted in average RMS error of 6.9 % over the training set and 13.9% error over the test set.
This chapter’s example in forecasting highlights the ease of use and wide applicability of the backpropagation
algorithm for large, complex problems and data sets. Several examples of research in financial forecasting
were presented with a number of ideas and real−life methodologies presented.
Technical Analysis was briefly discussed with examples of studies that can be useful in preprocessing data for
neural networks.
Summary 330
A Resource guide was presented for further information on financial applications of neural networks.
Summary 331

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Chapter 15
Application to Nonlinear Optimization
Introduction
Nonlinear optimization is an area of operations research, and efficient algorithms for some of the problems in
this area are hard to find. In this chapter, we describe the traveling salesperson problem and discuss how this
problem is formulated as a nonlinear optimization problem in order to use neural networks (Hopfield and
Kohonen) to find an optimum solution. We start with an explanation of the concepts of linear, integer linear
and nonlinear optimization.
An optimization problem has an objective function and a set of constraints on the variables. The problem is to
find the values of the variables that lead to an optimum value for the objective function, while satisfying all
the constraints. The objective function may be a linear function in the variables, or it may be a nonlinear
function. For example, it could be a function expressing the total cost of a particular production plan, or a
function giving the net profit from a group of products that share a given set of resources. The objective may
be to find the minimum value for the objective function, if, for example, it represents cost, or to find the
maximum value of a profit function. The resources shared by the products in their manufacturing are usually
in limited supply or have some other restrictions on their availability. This consideration leads to the
specification of the constraints for the problem.
Each constraint is usually in the form of an equation or an inequality. The left side of such an equation or
inequality is an expression in the variables for the problem, and the right−hand side is a constant. The
constraints are said to be linear or nonlinear depending on whether the expression on the left−hand side is a
linear function or nonlinear function of the variables. A linear programming problem is an optimization
problem with a linear objective function as well as a set of linear constraints. An integer linear programming
problem is a linear programming problem where the variables are required to have integer values. A nonlinear
optimization problem has one or more of the constraints nonlinear and/or the objective function is nonlinear.
Here are some examples of statements that specify objective functions and constraints:
• Linear objective function: Maximize Z = 3X 1 + 4X2 + 5.7X3

• Linear equality constraint: 13X 1 − 4.5X2 + 7X3 = 22
• Linear inequality constraint: 3.6X 1 + 8.4X2 − 1.7X3 d 10.9
• Nonlinear objective function: Minimize Z = 5X 2 + 7XY + Y2
• Nonlinear equality constraint: 4X + 3XY + 7Y + 2Y 2 = 37.6
• Nonlinear inequality constraint: 4.8X + 5.3XY + 6.2Y 2 e 34.56
An example of a linear programming problem is the blending problem. An example of a blending problem is
that of making different flavors of ice cream blending different ingredients, such as sugar, a variety of nuts,
and so on, to produce different amounts of ice cream of many flavors. The objective in the problem is to find
Chapter 15 Application to Nonlinear Optimization 332

the amounts of individual flavors of ice cream to produce with given supplies of all the ingredients, so the
total profit is maximized.
A nonlinear optimization problem example is the quadratic programming problem. The constraints are all
linear but the objective function is a quadratic form. A quadratic form is an expression of two variables with 2
for the sum of the exponents of the two variables in each term.
An example of a quadratic programming problem, is a simple investment strategy problem that can be stated
as follows. You want to invest a certain amount in a growth stock and in a speculative stock, achieving at least
25% return. You want to limit your investment in the speculative stock to no more than 40% of the total
investment. You figure that the expected return on the growth stock is 18%, while that on the speculative
stock is 38%. Suppose G and S represent the proportion of your investment in the growth stock and the
speculative stock, respectively. So far you have specified the following constraints. These are linear
constraints:
G+S=1 This says the proportions add up to 1.

S d 0.4 This says the proportion invested in speculative stock is no more than 40%.
1.18G + 1.38S e 1.25 This says the expected return from these investments should be at least 25%.
Now the objective function needs to be specified. You have specified already the expected return you want to
achieve. Suppose that you are a conservative investor and want to minimize the variance of the return. The
variance works out as a quadratic form. Suppose it is determined to be:
2G2 + 3S2 − GS
This quadratic form, which is a function of G and S, is your objective function that you want to minimize
subject to the (linear) constraints previously stated.
Neural Networks for Optimization Problems

It is possible to construct a neural network to find the values of the variables that correspond to an optimum
value of the objective function of a problem. For example, the neural networks that use the Widrow−Hoff
learning rule find the minimum value of the error function using the least mean squared error. Neural
networks such as the feedforward backpropagation network use the steepest descent method for this purpose
and find a local minimum of the error, if not the global minimum. On the other hand, the Boltzmann machine
or the Cauchy machine uses statistical methods and probabilities and achieves success in finding the global
minimum of an error function. So we have an idea of how to go about using a neural network to find an
optimum value of a function. The question remains as to how the constraints of an optimization problem
should be treated in a neural network operation. A good example in answer to this question is the traveling
salesperson problem. Let’s discuss this example next.
Neural Networks for Optimization Problems 333


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Traveling Salesperson Problem

The traveling salesperson problem is well−known in optimization. Its mathematical formulation is simple, and
one can state a simple solution strategy also. Such a strategy is often impractical, and as yet, there is no
efficient algorithm for this problem that consistently works in all instances. The traveling salesperson problem
is one among the so− called NP−complete problems, about which you will read more in what follows. That
means that any algorithm for this problem is going to be impractical with certain examples. The neural
network approach tends to give solutions with less computing time than other available algorithms for use on
a digital computer. The problem is defined as follows. A traveling salesperson has a number of cities to visit.
The sequence in which the salesperson visits different cities is called a tour. A tour should be such that every
city on the list is visited once and only once, except that he returns to the city from which he starts. The goal is
to find a tour that minimizes the total distance the salesperson travels, among all the tours that satisfy this
criterion.
A simple strategy for this problem is to enumerate all feasible tours—a tour is feasible if it satisfies the
criterion that every city is visited but once—to calculate the total distance for each tour, and to pick the tour
with the smallest total distance. This simple strategy becomes impractical if the number of cities is large. For
example, if there are 10 cities for the traveling salesperson to visit (not counting home), there are 10! =
3,628,800 possible tours, where 10! denotes the factorial of 10—the product of all the integers from 1 to
10—and is the number of distinct permutations of the 10 cities. This number grows to over 6.2 billion with
only 13 cities in the tour, and to over a trillion with 15 cities.
The TSP in a Nutshell
For n cities to visit, let Xij be the variable that has value 1 if the salesperson goes from city i to city j and value
0 if the salesperson does not go from city i to city j. Let dij be the distance from city i to city j. The traveling
salesperson problem (TSP) is stated as follows:
Minimize the linear objective function:
subject to:
for each j=1, …, n (linear constraint)
Traveling Salesperson Problem 334

for each i=1, …, n (linear constraint)
wXij for 1 for all i and j (integer constraint)
This is a 0−1 integer linear programming problem.
Solution via Neural Network
This section shows how the linear and integer constraints of the TSP are absorbed into an objective function
that is nonlinear for solution via Neural network.
The first consideration in the formulation of an optimization problem is the identification of the underlying
variables and the type of values they can have. In a traveling salesperson problem, each city has to be visited
once and only once, except the city started from. Suppose you take it for granted that the last leg of the tour is
the travel between the last city visited and the city from which the tour starts, so that this part of the tour need
not be explicitly included in the formulation. Then with n cities to be visited, the only information needed for
any city is the position of that city in the order of visiting cities in the tour. This suggests that an ordered
n−tuple is associated with each city with some element equal to 1, and the rest of the n – 1 elements equal to
0. In a neural network representation, this requires n neurons associated with one city. Only one of these n
neurons corresponding to the position of the city, in the order of cities in the tour, fires or has output 1. Since
there are n cities to be visited, you need n2 neurons in the network. If these neurons are all arranged in a
square array, you need a single 1 in each row and in each column of this array to indicate that each city is
visited but only once.
Let xij be the variable to denote the fact that city i is the jth city visited in a tour. Then xij is the output of the
jth neuron in the array of neurons corresponding to the ith city. You have n2 such variables, and their values
are binary, 0 or 1. In addition, only n of these variables should have value 1 in the solution. Furthermore,
exactly one of the x’s with the same first subscript (value of i) should have value 1. It is because a given city
can occupy only one position in the order of the tour. Similarly, exactly one of the x’s with the same second
subscript (value of j) should have value 1. It is because a given position in the tour can be only occupied by
one city. These are the constraints in the problem. How do you then describe the tour? We take as the starting
city for the tour to be city 1 in the array of cities. A tour can be given by the sequence 1, a, b, c, …, q,
indicating that the cities visited in the tour in order starting at 1 are, a, b, c, …, q and back to 1. Note that the
sequence of subscripts a, b, …, q is a permutation of 2, 3, … n – 1, xa1=1, xb2=1, etc.
Having frozen city 1 as the first city of the tour, and noting that distances are symmetric, the distinct number
of tours that satisfy the constraints is not n!, when there are n cities in the tour as given earlier. It is much less,
namely, n!/2n. Thus when n is 10, the number of distinct feasible tours is 10!/20, which is 181,440. If n is 15,
it is still over 43 billion, and it exceeds a trillion with 17 cities in the tour. Yet for practical purposes there is
not much comfort knowing that for the case of 13 cities, 13! is over 6.2 billion and 13!/26 is only 239.5
million—it is still a tough combinatorial problem.
Solution via Neural Network 335


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
The cost of the tour is the total distance traveled, and it is to be minimized. The total distance traveled in the
tour is the sum of the distances from each city to the next. The objective function has one term that
corresponds to the total distance traveled in the tour. The other terms, one for each constraint, in the objective
function are expressions, each attaining a minimum value if and only if the corresponding constraint is
satisfied. The objective function then takes the following form. Hopfield and Tank formulated the problem as
one of minimizing energy. Thus it is, customary to refer to the value of the objective function of this problem,
while using Hopfield−like neural network for its solution, as the energy level, E, of the network. The goal is to
minimize this energy level.
In formulating the equation for E, one uses constant parameters A1, A2, A3, and A4 as coefficients in different
terms of the expression on the right−hand side of the equation. The equation that defines E is given as follows.
Note that the notation in this equation includes dij for the distance from city i to city j.
E = A1 £i£k £j`kxikxij + A2 £i £k xi£k £j`kxkixji + A3[( £i£kxik) − n]2 + A4£k £k £j`k £idkjxki(xj,i+1 + xj,i−1)
Our first observation at this point is that E is a nonlinear function of the x’s, as you have quadratic terms in it.
So this formulation of the traveling salesperson problem renders it a nonlinear optimization problem.
All the summations indicated by the occurrences of the summation symbol £, range from 1 to n for the values
of their respective indices. This means that the same summand such as x12x33 also as x33x12, appears twice
with only the factors interchanged in their order of occurrence in the summand. For this reason, many authors
use an additional factor of 1/2 for each term in the expression for E. However, when you minimize a quantity
z with a given set of values for the variables in the expression for z, the same values for these variables
minimize any whole or fractional multiple of z, as well.
The third summation in the first term is over the index j, from 1 to n, but excluding whatever value k has. This
prevents you from using something like x12x12. Thus, the first term is an abbreviation for the sum of n2(n – 1)
terms with no two factors in a summand equal. This term is included to correspond to the constraint that no
more than one neuron in the same row can output a 1. Thus, you get 0 for this term with a valid solution. This
is also true for the second term in the right−hand side of the equation for E. Note that for any value of the
index i, xii has value 0, since you are not making a move like, from city i to the same city i in any of the tours
you consider as a solution to this problem. The third term in the expression for E has a minimum value of 0,
which is attained if and only if exactly n of the n2 x’s have value 1 and the rest 0.
The last term expresses the goal of finding a tour with the least total distance traveled, indicating the shortest
tour among all possible tours for the traveling salesperson. Another important issue about the values of the
subscripts on the right−hand side of the equation for E is, what happens to i + 1, for example, when i is
already equal to n, and to i–1, when i is equal to 1. The i + 1 and i – 1 seem like impossible values, being out
of their allowed range from 1 to n. The trick is to replace these values with their moduli with respect to n. This
means, that the value n + 1 is replaced with 1, and the value 0 is replaced with n in the situations just
described.

Modular values are obtained as follows. If we want, say 13 modulo 5, we subtract 5 as many times as possible
from 13, until the remainder is a value between 0 and 4, 4 being 5 – 1. Since we can subtract 5 twice from 13
to get a remainder of 3, which is between 0 and 4, 3 is the value of 13 modulo 5. Thus (n + 3) modulo n is 3,
as previously noted. Another way of looking at these results is that 3 is 13 modulo 5 because, if you subtract 3
from 13, you get a number divisible by 5, or which has 5 as a factor. Subtracting 3 from n + 3 gives you n,
which has n as a factor. So 3 is (n + 3) modulo n. In the case of –1, by subtracting (n – 1) from it, we get −n,
which can be divided by n getting –1. So (n – 1) is the value of (–1) modulo n.


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Example of a Traveling Salesperson Problem for Hand Calculation
Suppose there are four cities in the tour. Call these cities, C1, C2, C3, and C4. Let the matrix of distances be the
following matrix D.
0 10 14 7
D = 10 0 6 12
14 6 0 9
7 12 9 0
From our earlier discussion on the number of valid and distinct tours, we infer that there are just three such
tours. Since it is such a small number, we can afford to enumerate the three tours, find the energy values
associated with them, and pick the tour that has the smallest energy level among the three. The three tours are:
• Tour 1. 1 − 2 − 3 − 4 − 1 In this tour city 2 is visited first, followed by city 3, from where the
salesperson goes to city 4, and then returns to city 1. For city 2, the corresponding ordered array is (1,
0, 0, 0), because city 2 is the first in this permutation of cities. Then x21 = 1, x22 = 0, x23 = 0, x24 = 0.
Also (0, 1, 0, 0), (0, 0, 1, 0), and (0, 0, 0, 1) correspond to cities 3, 4, and 1, respectively. The total
distance of the tour is d12 + d23 + d34 + d41= 10 + 6 + 9 + 7 = 32.
• Tour 2. 1 − 3 − 4 − 2 − 1
• Tour 3. 1 − 4 − 2 − 3 − 1
There seems to be some discrepancy here. If there is one, we need an explanation. The discrepancy is that we
can find many more tours that should be valid because no city is visited more than once. You may say they are
distinct from the three previously listed. Some of these additional tours are:
• Tour 4. 1 − 2 − 4 − 3 − 1
• Tour 5. 3 − 2 − 4 − 1 − 3
• Tour 6. 2 − 1 − 4 − 3 − 2
There is no doubt that these three tours are distinct from the first set of three tours. And in each of these three
tours, every city is visited exactly once, as required in the problem. So they are valid tours as well. Why did
our formula give us 3 for the value of the number of possible valid tours, while we are able to find 6?
The answer lies in the fact that if two valid tours are symmetric and have the same energy level, because they
have the same value for the total distance traveled in the tour, one of them is in a sense redundant, or one of
them can be considered degenerate, using the terminology common to this context. As long as they are valid
and give the same total distance, the two tours are not individually interesting, and any one of the two is
enough to have. By simple inspection, you find the total distances for the six listed tours. They are 32 for tour
1, 32 also for tour 6, 45 for each of tours 2 and 4, and 39 for each of tours 3 and 5. Notice also that tour 6 is
not very different from tour 1. Instead of starting at city 1 as in tour 1, if you start at city 2 and follow tour 1
from there in reverse order of visiting cities, you get tour 6. Therefore, the distance covered is the same for
both these tours. You can find similar relationships for other pairs of tours that have the same total distance
Example of a Traveling Salesperson Problem for Hand Calculation 338

for the tour. Either by reversing the order of cities in a tour, or by making a circular permutation of the order
of the cities in a tour, you get another tour with the same total distance. This way you can find tours. Thus,
only three distinct total distances are possible, and among them 32 is the lowest. The tour 1 − 2 − 3 − 4 − 1 is
the shortest and is an optimum solution for this traveling salesperson problem. There is an alternative
optimum solution, namely tour 6, with 32 for the total length of the tour. The problem is to find an optimal
tour, and not to find all optimal tours if more than one exist. That is the reason only three distinct tours are
suggested by the formula for the number of distinct valid tours in this case.
The formula we used to get the number of valid and distinct tours to be 3 is based on the elimination of such
symmetry. To clarify all this discussion, you should determine the energy levels of all your six tours identified
earlier, hoping to find pairs of tours with identical energy levels.
Note that the first term on the right−hand side of the equation results in 0 for a valid tour, as this term is to
ensure there is no more than a single 1 in each row. That is, in any summand in the first term, at least one of
the factors, xik or xij, where k ` j has to be 0 for a valid tour. So all those summands are 0, making the first term
itself 0. Similarly, the second term is 0 for a valid tour, because in any summand at least one of the factors, xki
or xji, where k ` j has to be 0 for a valid tour. In all, exactly 4 of the 16 x’s are each 1, making the total of the
x’s 4. This causes the third term to be 0 for a valid tour. These observations make it clear that it does not
matter for valid tours, what values are assigned to the parameters A1, A2, and A3. Assigning large values for
these parameters would cause the energy levels, for tours that are not valid, to be much larger than the energy
levels for valid tours. Thereby, these tours become unattractive for the solution of the traveling salesperson
problem. Let us use the value 1/2 for the parameter A4.
Let us demonstrate the calculation of the value for the last term in the equation for E, in the case of tour 1.
Recall that the needed equation is
E = A1 £i£k £j`k xikxij + A2 £i£k £j`kxkixji + A3[( £i£kxjk) − n]2 + A4£k £j`k£idkjxki(xj,i+1 + xj,i−1)
The last term expands as given in the following calculation:
A4{ d12[x12(x23 +x21) + x13(x24 + x22) + x14(x21 + x23)] +

d13[x12(x33 + x31) + x13(x34 + x32) + x14(x31 + x33)] +
d14[x12(x43 + x41) + x13(x44 + x42) + x14(x41 + x43)] +
d21[x21(x12 + x14) + x23(x14 + x12) + x24(x11 + x13)] +
d23[x21(x32 + x34) + x23(x34 + x32) + x24(x31 + x33)] +
d24[x21(x42 + x44) + x23(x44 + x42) + x24(x41 + x43)] +
d31[x31(x12 + x14) + x32(x13 + x11) + x34(x11 + x13)] +
d32[x31(x22 + x24) + x32(x23 + x21) + x34(x21 + x23)] +
d34[x31(x42 + x44) + x32(x43 + x41) + x34(x41 + x43)] +
d41[x41(x12 + x14) + x42(x13 + x11) + x43(x14 + x12)] +
d42[x41(x22 + x24) + x42(x23 + x21) + x43(x24 + x22)] +
d43[x41(x32 + x34) + x42(x33 + x31) + x43(x34 + x32)] }
When the respective values are substituted for the tour 1 − 2 − 3 − 4 − 1, the previous calculation becomes:
1/2{10[0(0 + 1) + 0(0 + 0) + 1(1 + 0)] + 14[0(0 + 0) + 0(0 + 1) +

1(0 + 0)] +
7[0(1 + 0) + 0(0 + 0) + 1(0 + 1)] + 10[1(0 + 1) + 0(1 + 0) +
0(0 + 0)] +
6[1(1 + 0) + 0(0 + 1) + 0(0 + 0)] + 12[1(0 + 0) + 0(0 + 0) +
0(0 + 1)] +
14[0(0 + 1) + 1(0 + 0) + 0(0 + 0)] + 6[0(0 + 0) + 1(0 + 1) +
0(1 + 0)] +
9[0(0 + 0) + 1(1 + 0) + 0(0 + 1)] + 7[0(0 + 1) + 0(0 + 0) +
1(1 + 0)] +

12[0(0 + 0) + 0(0 + 1) + 1(0 + 0)] + 9[0(1 + 0) + 0(0 + 0) +

1(0 + 1)]}
= 1/2( 10 + 0 + 7 + 10 + 6 + 0 + 0 + 6 + 9 + 7 + 0 + 9)
= 1/2(64)
= 32
Table 15.1 contains the values we get for the fourth term on the right−hand side of the equation, and for E,
with the six listed tours.
Table 15.1 Energy Levels for Six of the Valid Tours

Tour #Non−Zero x’sValue for the Last TermEnergy LevelComment
1x14, x21, x32, x4332321 − 2 − 3 − 4 − 1 tour
2x14, x23, x31, x4245451 − 3 − 4 − 2 − 1 tour
3x14, x22, x33, x4139391 − 4 − 2 − 3 − 1 tour
4x14, x21, x33, x4245451 − 2 − 4 − 3 − 1 tour
5x13, x21, x34, x4239393 − 2 − 4 − 1 − 3 tour
6x11, x24, x33, x4232322 − 1 − 4 − 3 − 2 tour


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Neural Network for Traveling Salesperson Problem

Hopfield and Tank used a neural network to solve a traveling salesperson problem. The solution you get may
not be optimal in certain instances. But by and large you may get a solution close to an optimal solution. One
cannot find for the traveling salesperson problem a consistently efficient method of solution, as it has been
proved to belong to a set of problems called NP−complete problems. NP−complete problems have the
distinction that there is no known algorithm that is efficient and practical, and there is little likelihood that
such an algorithm will be developed in the future. This is a caveat to keep in mind when using a neural
network to solve a traveling salesperson problem.
Network Choice and Layout
We will describe the use of a Hopfield network to attempt to solve the traveling salesperson problem. There
are n2 neurons in the network arranged in a two−dimensional array of n neurons per row and n per column.
The network is fully connected. The connections in the network in each row and in each column are lateral
connections. The layout of the neurons in the network with their connections is shown in Figure 15.1 for the
case of three cities, for illustration. To avoid cluttering, the connections between diagonally opposite neurons
are not shown.
Figure 15.1 Layout of a Hopfield network for the traveling salesperson problem.
The most important task on hand then is finding an appropriate connection weight matrix. It is constructed
taking into account that nonvalid tours should be prevented and valid tours should be preferred. A
consideration in this regard is, for example, no two neurons in the same row (column) should fire in the same
cycle of operation of the network. As a consequence, the lateral connections should be for inhibition and not
for excitation.
In this context, the Kronecker delta function is used to facilitate simple notation. The Kronecker delta
function has two arguments, which are given usually as subscripts of the symbol ´. By definition ík has value
1 if i = k, and 0 if i ` k. That is, the two subscripts should agree for the Kronecker delta to have value 1.
Otherwise, its value is 0.
We refer to the neurons with two subscripts, one for the city it refers to, and the other for the order of that city
in the tour. Therefore, an element of the weight matrix for a connection between two neurons needs to have
four subscripts, with a comma after two of the subscripts. An example is wik,lj referring to the weight on the
connection between the (ik) neuron and the (lj) neuron. The value of this weight is set as follows:
Wik,lj = −A1íl(1−´kj)−A2´kj(1−´kj(1−íl)−A3−A4 dil(´j, k+1 + ´j,k−1)
Neural Network for Traveling Salesperson Problem 341

Here the negative signs indicate inhibition through the lateral connections in a row or column. The −A3 is a
term for global inhibition.
Inputs
The inputs to the network are chosen arbitrarily. If as a consequence of the choice of the inputs, the
activations work out to give outputs that add up to the number of cities, an initial solution for the problem, a
legal tour, will result. A problem may also arise that the network will get stuck at a local minimum. To avoid
such an occurrence, random noise can be added. Usually you take as the input at each neuron a constant times
the number of cities and adjust this adding a random number, which can differ for different neurons.
Activations, Outputs, and Their Updating
We denote the activation of the neuron in the ith row and jth column by aij, and the output is denoted by xij. A
time constant Ä, and a gain » are used as well. A constant m is another parameter used. Also, ”t denotes the
increment in time, from one cycle to the next. Keep in mind that the index for the summation £ ranges from 1
to n, the number of cities. Excluded values of the index are shown by the use of the symbol `.
The change in the activation is then given by ”aij, where:
”aij = ”t (Term1 + Term2 + Term3 + Term4 + Term5)

Term1 = − aij/Ä
Term2 = − A1 k`jxik
Term3 = − A2
£kìxkj
Term4 = − A3(£i £kxik − m)
Term5 = − A4 £kìdik(xk,j+1 + xk,j−1)
To update the activation of the neuron in the ith row and jth column, you take:
a
ijnew = aijold + ”aij
The output of a neuron in the ith row and jth column is calculated from:
xin = (1 + tanh(»aij))/2
NOTE:
, which is the original sigmoid function
The function used here is the hyperbolic tangent function. The gain parameter mentioned earlier » is. The
output of each neuron is calculated after updating the activation. Ideally, you want to get the outputs as 0’s
and 1’s, preferably a single one for each row and each column, to represent a tour that satisfies the conditions
of the problem. But the hyperbolic tangent function gives a real number, and you have to settle for a close
enough value to 1 or 0. You may get, for example, 0.96 instead of 1, or 0.07 instead of 0. The solution is to be
obtained from such values by rounding up or down so that 1 or 0 will be used, as the case may be
Inputs 342
Inputs 343

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Performance of the Hopfield Network
Let us now look at Hopfield and Tank’s approach at solving the TSP.
Hopfield and Tank Example

The Hopfield network’s use in solving the traveling salesperson problem is a pioneering effort in the use of
the neural network approach for this problem. Hopfield and Tank’s example is for a problem with 10 cities.
The parameters used were, A1= 500, A2 = 500, A3 = 200, A4 = 500, Ä = 1, » = 50, and m = 15. A good solution
corresponding to a local minimum for E is the expected, if not the best, solution (global minimum). An
annealing process could be considered to move out of any local minimum. As was mentioned before, the
traveling salesperson problem is one of those problems for which a single approach cannot be found that will
be successful in all cases. There isn’t very much guidance as to how to choose the parameters in general for
the use of the Hopfield network to solve the traveling salesperson problem.
C++ Implementation of the Hopfield Network for the Traveling Salesperson Problem
We present a C++ program for the Hopfield network operation for the traveling salesperson problem. The
header file is in the Listing 15.1, and the source file is in the Listing 15.2. A tsneuron class is declared for a
neuron and a network class for the network. The network class is declared a friend in the tsneuron class.
The program follows the procedure described for setting inputs, connection weights, and updating.
Program Details
The following is a listing of the characteristics of the C++ program along with definitions and/or functions.
• The number of cities, the number of iterations, and the distances between the cities are solicited
from the user.
• The distances are taken as integer values. If you want to use real numbers as distances, the type
for distance matrix needs to be changed to float, and corresponding changes are needed for calcdist (
) function, etc.
• tourcity and tourorder arrays keep track of the cities that have to be covered and the order in
which it is to be done.
• A neuron corresponds to each combination of a city and its order in the tour. The ith city visited in
the order j , is the neuron corresponding to the element j + i*n, in the array for neurons. Here n is the
number of cities. The i and the j vary from 0 to n – 1. There are n2 neurons.
• mtrx is the matrix giving the weights on the connections between the neurons. It is a square
matrix of order n2.
• An input vector is generated at random in the function main ( ), and is later referred to as ip
• asgninpt ( ) function presents the input vector ip to the network and determines the initial
activations of the neurons.
• getacts ( ) function updates the activations of the neurons after each iteration.
Performance of the Hopfield Network 344

• getouts ( ) function computes the outputs after each iteration. la is used as abbreviation for
lambda in its argument.
• iterate ( ) function carries out the number of iterations desired.
• findtour ( ) function determines the tour orders of cities to be visited using the outputs of the
neurons. When used at the end of the iterations, it gives the solution obtained for the traveling
salesperson problem.
• calcdist ( ) function calculates the distance of the tour in the solution.
Listing 15.1 Header file for the C++ program for the Hopfield network for the traveling salesperson problem
//trvslsmn.h V. Rao, H. Rao

#include <stdlib.h>
#include <math.h>
#include <stdio.h>
#define MXSIZ 11
class tsneuron
{
protected:
int cit,ord;
float output;
float activation;
public:
tsneuron() { };
void getnrn(int,int);
};
class network
{
public:
int citnbr;
float pra,prb,prc,prd,totout,distnce;
tsneuron (tnrn)[MXSIZ][MXSIZ];
int dist[MXSIZ][MXSIZ];
int tourcity[MXSIZ];
int tourorder[MXSIZ];
float outs[MXSIZ][MXSIZ];
float acts[MXSIZ][MXSIZ];
float mtrx[MXSIZ][MXSIZ];
float citouts[MXSIZ];
float ordouts[MXSIZ];
network() { };
void getnwk(int,float,float,float,float);
void getdist(int);
void findtour();
void asgninpt(float *);
void calcdist();
void iterate(int,int,float,float,float);
void getacts(int,float,float);
void getouts(float);
//print functions
void prdist();

void prmtrx(int);
void prtour();
void practs();
void prouts();
};


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Source File for Hopfield Network for Traveling Salesperson Problem
The following listing gives the source code for the C++ program for the Hopfield network for traveling
salesperson problem. The user is prompted to input the number of cities and the maximum number of
iterations for the operation of the network.
The parameters a, b, c, d declared in the function main correspond to A1, A2, A3, and A4, respectively. These
and the parameters tau, lambda, and nprm are all given values in the declaration statements in the function
main. If you change these parameter values or change the number of cities in the traveling salesperson
problem, the program will compile but may not work as you’d like.
Listing 15.2 Source file for the C++ program for the Hopfield network for the traveling salesperson problem
//trvslsmn.cpp V. Rao, H. Rao

#include “trvslsmn.h”
#include <stdlib.h>
#include <time.h>
//generate random noise
int randomnum(int maxval)

{
// will return an integer up to maxval
return rand() % maxval;

}
//Kronecker delta function
int krondelt(int i,int j)

{
int k;
k= ((i == j) ? (1):(0));
return k;
};
void tsneuron::getnrn(int i,int j)

{
cit = i;
ord = j;
output = 0.0;
activation = 0.0;
};
//distances between cities
void network::getdist(int k)
{
Source File for Hopfield Network for Traveling Salesperson Problem 347
citnbr = k;
int i,j;
cout<<”\n”;
for(i=0;i<citnbr;++i)
{
dist[i][i]=0;
for(j=i+1;j<citnbr;++j)
{
cout<<”\ntype distance (integer) from city “<<
i<<” to city “<<j<<”\n”;
cin>>dist[i][j];
}
cout<<”\n”;
}
{
for(j=0;j<i;++j)
{
dist[i][j] = dist[j][i];
}
}
prdist();
cout<<”\n”;
}
//print distance matrix
void network::prdist()
{
int i,j;
cout<<”\n Distance Matrix\n”;
{
for(j=0;j<citnbr;++j)
{
cout<<dist[i][j]<<” “;
}
cout<<”\n”;
}
}
//set up network
void network::getnwk(int citynum,float a,float b,float c,float d)

{
int i,j,k,l,t1,t2,t3,t4,t5,t6;
int p,q;
citnbr = citynum;
pra = a;
prb = b;
prc = c;
prd = d;
getdist(citnbr);
{
{
tnrn[i][j].getnrn(i,j);
}
}
//find weight matrix
{
{
p = ((j == citnbr−1) ? (0) : (j+1));
q = ((j == 0) ? (citnbr−1) : (j−1));
t1 = j + i*citnbr;
for(k=0;k<citnbr;++k)
{
for(l=0;l<citnbr;++l)
{
t2 = l + k*citnbr;
t3 = krondelt(i,k);
t4 = krondelt(j,l);
t5 = krondelt(l,p);
t6 = krondelt(l,q);
mtrx[t1][t2] =
−a*t3*(1−t4) −b*t4*(1−t3)
−c −d*dist[i][k]*(t5+t6);
}
}
}
}
prmtrx(citnbr);
}
//print weight matrix
void network::prmtrx(int k)
{
int i,j,nbrsq;
nbrsq = k*k;
cout<<”\nWeight Matrix\n”;
for(i=0;i<nbrsq;++i)
{
for(j=0;j<nbrsq;++j)
{
if(j%k == 0)
{
cout<<”\n”;
}
cout<<mtrx[i][j]<<” “;
}
cout<<”\n”;
}
}
//present input to network
void network::asgninpt(float *ip)

{
int i,j,k,l,t1,t2;
{
for(j =0;j<citnbr;++j)
{
acts[i][j] = 0.0;
}
}
//find initial activations

{
for(j =0;j<citnbr;++j)
{
t1 = j + i*citnbr;
{
for(l=0;l<citnbr;++l)
{
t2 = l + k*citnbr;
acts[i][j] +=
mtrx[t1][t2]*ip[t1];
}
}
}
}
//print activations
cout<<”\ninitial activations\n”;
practs();
}
//find activations
void network::getacts(int nprm,float dlt,float tau)

{
int i,j,k,p,q;
float r1, r2, r3, r4,r5;
r3 = totout − nprm ;
{
r4 = 0.0;
p = ((i == citnbr−1) ? (0) : (i+1));
q = ((i == 0) ? (citnbr−1) : (i−1));
{
r1 = citouts[i] − outs[i][j];
r2 = ordouts[i] − outs[i][j];
{
r4 += dist[i][k] *
(outs[k][p] + outs[k][q]);
}
r5 = dlt*(−acts[i][j]/tau −
pra*r1 −prb*r2 −prc*r3 −prd*r4);
acts[i][j] += r5;
}
}
}
//find outputs and totals for rows and columns
void network::getouts(float la)

{
float b1,b2,b3,b4;
int i,j;
totout = 0.0;
{
citouts[i] = 0.0;
{
b1 = la*acts[i][j];
b4 = b1/500.0;
b2 = exp(b4);
b3 = exp(−b4);
outs[i][j] = (1.0+(b2−b3)/(b2+b3))/2.0;
citouts[i] += outs[i][j];};
totout += citouts[i];
}
{
ordouts[j] = 0.0;
{
ordouts[j] += outs[i][j];
}
}
}
//find tour
void network::findtour()
{
int i,j,k,tag[MXSIZ][MXSIZ];
float tmp;
for (i=0;i<citnbr;++i)
{
{
tag[i][j] = 0;
}
}
for (i=0;i<citnbr;++i)
{
tmp = −10.0;
{
{
if((outs[i][k] >=tmp)&&
(tag[i][k] ==0))
tmp = outs[i][k];
}
if((outs[i][j] ==tmp)&&
(tag[i][j]==0))
{
tourcity[i] =j;
tourorder[j] = i;
cout<<”\ntourcity “<<i
<<” tour order “<<j<<”\n”;
{
tag[i][k] = 1;
tag[k][j] = 1;
}
}
}
}
}
//print outputs
void network::prouts()
{
int i,j;
cout<<”\nthe outputs\n”;
{
{
cout<<outs[i][j]<<” “;
}
cout<<”\n”;
}
}
//calculate total distance for tour
void network::calcdist()
{
int i, k, l;
distnce = 0.0;
{
k = tourorder[i];
l = ((i == citnbr−1 ) ? (tourorder[0]):(tourorder[i+1]));
distnce += dist[k][l];
}
cout<<”\n distance of tour is : “<<distnce<<”\n”;
}
// print tour
void network::prtour()
{
int i;
cout<<”\n the tour :\n”;
{
cout<<tourorder[i]<<” “;
cout<<”\n”;
}
}
//print activations
void network::practs()
{
int i,j;
cout<<”\n the activations:\n”;
{
{
cout<<acts[i][j]<<” “;
}
cout<<”\n”;
}
}
//iterate specified number of times
void network::iterate(int nit,int nprm,float dlt,float tau,float la)

{
int k;
for(k=1;k<=nit;++k)
{
getacts(nprm,dlt,tau);
getouts(la);
}
cout<<”\n” <<nit<<” iterations completed\n”;
practs();
cout<<”\n”;
prouts();
cout<<”\n”;
}
void main()
{
//numit = #iterations; n = #cities; u=intial input; nprm − parameter n’

//dt = delta t;
// —————————————————−
// parameters to be tuned are here:
int u=1;
int nprm=10;
float a=40.0;
float b=40.0;
float c=30.0;
float d=60.0;
float dt=0.01;
float tau=1.0;
float lambda=3.0;
//——————————————————−
int i,n2;
int numit=100;
int n=4;
float input_vector[MXSIZ*MXSIZ];
cout<<”\nPlease type number of cities, number of iterations\n”;
cin>>n>>numit;
cout<<”\n”;
if (n>MXSIZ)
{
cout << “choose a smaller n value\n”;
exit(1);
}
n2 = n*n;
for(i=0;i<n2;++i)
{
if(i%n == 0)cout<<”\n”;
input_vector[i] =(u + (float)(randomnum(100)/100.0))/20.0;
cout<<input_vector[i]<<” “;
}
//create network and operate
network *tntwk = new network;

if (tntwk==0)
{
cout << “not enough memory\n”;
exit(1);
}
tntwk−>getnwk(n,a,b,c,d);
tntwk−>asgninpt(input_vector);
tntwk−>getouts(lambda);
tntwk−>prouts();
tntwk−>iterate(numit,nprm,dt,tau,lambda);
tntwk−>findtour();
tntwk−>prtour();
tntwk−>calcdist();
cout<<”\n”;
}

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Output from Your C++ Program for the Traveling Salesperson Problem
A three−city tour problem is trivial, since there is just one value for the total distance no matter how you
permute the cities for the order of the tour. In this case the natural order is itself an optimal solution. The
program is run for two cases, for illustration. The first run is for a problem with four cities. The second one is
for a five−city problem. By the way, the cities are numbered from 0 to n – 1. The same parameter values are
used in the two runs. The number of cities, and consequently, the matrix of distances were different. In the
first run, the number of iterations asked for is 30, and in the second run it is 40.
The solution you get for the four−city problem is not the one in natural order. The total distance of the tour is
32. The tour in the solution is 1 − 0 − 3 − 2 − 1. This tour is equivalent to the tour in natural order, as you can
see by starting at 0 and reading cyclically right to left in the previous sequence of cities.
The solution you get for the five−city problem is the tour 1 − 2 − 0 − 4 − 3 − 1. This reads, starting from 0 as
either 0 − 4 − 3 − 1 − 2 − 0, or as 0 − 2 − 1 − 3 − 4 − 0. It is different from the tour in natural order. It has a
total distance of 73, as compared to the distance of 84 with the natural order. It is not optimal though, as a tour
of shortest distance is 0 − 4 − 2 − 1 − 3 − 0 with total distance 50.
Can the program give you the shorter tour? Yes, the solution can be different if you run the program again
because each time you run the program, the input vector is different, as it is chosen at random. The parameter
values given in the program are by guess.
Note that the seed for the random number generator is given in the statement
The program gives you the order in the tour for each city. For example, if it says tourcity 1 tour order 2, that
means the second (tour) city is the city visited third (in tour order). Your tour orders are also with values from
0 to n – 1, like the cities.
The user input is in italic, and computer output is normal, as you have seen before.
Output for a Four−City Problem

Please type number of cities, number of iterations
4 30
0.097 0.0585 0.053 0.078 //input vector—there

are 16 neurons in the network
0.0725 0.0535 0.0585 0.0985
0.0505 0.061 0.0735 0.057
0.0555 0.075 0.0885 0.0925
type distance (integer) from city 0 to city 1
Output from Your C++ Program for the Traveling Salesperson Problem 355
10
14
7

6
12

9
Distance Matrix
0 10 14 7
10 0 6 12
14 6 0 9
7 12 9 0
Weight Matrix //16x16 matrix of weights. There are

16 neurons in the network.
−30 −70 −70 −70
−70 −630 −30 −630
−70 −870 −30 −70
−30 −70 −70 −630
−70 −30 −70 −70

−630 −70 −630 −30
−870 −70 −870 −70
−70 −30 −70 −30
−70 −70 −30 −70

−30 −630 −70 −630
−30 −870 −70 −70
−70 −70 −30 −630
−70 −70 −70 −30

−630 −30 −630 −70
−870 −30 −870 −70
−630 −30 −630 −30
−70 −630 −30 −630

−30 −70 −70 −70
−70 −390 −30 −630
−70 −630 −30 −70
−630 −70 −630 −30

−70 −30 −70 −70
−390 −70 −390 −30
−630 −70 −630 −70
−30 −630 −70 −630

−70 −70 −30 −70
−30 −390 −70 −630
−30 −630 −70 −70
−630 −30 −630 −70

−70 −70 −70 −30
−390 −30 −390 −70
−870 −30 −870 −70
−70 −870 −30 −870

−70 −390 −30 −390
−30 −70 −70 −870
−70 −870 −30 −390
−870 −70 −870 −30

−390 −70 −390 −30
−70 −30 −70 −30
−870 −70 −870 −30
−30 −870 −70 −870

−30 −390 −70 −390
−70 −70 −30 −870
−30 −870 −70 −390
−870 −30 −870 −70

−390 −30 −390 −70
−70 −70 −70 −70
−450 −30 −450 −70
−70 −450 −30 −450

−70 −750 −30 −750
−70 −570 −30 −450
−70 −450 −30 −750
−450 −70 −450 −30

−750 −70 −750 −30
−570 −70 −570 −30
−450 −70 −450 −30
−30 −450 −70 −450

−30 −750 −70 −750
−30 −570 −70 −450
−30 −450 −70 −750
−450 −30 −450 −70

−750 −30 −750 −70
−570 −30 −570 −70
−70 −70 −70 −30
initial activations
the activations:
−333.680054 −215.280014 −182.320023 −371.280029
−255.199997 −207.580002 −205.920013 −425.519989
−258.560028 −290.360016 −376.320007 −228.000031
−278.609985 −363 −444.27005 −377.400024
the outputs
0.017913 0.070217 0.100848 0.011483
0.044685 0.076494 0.077913 0.006022
0.042995 0.029762 0.010816 0.060882
0.034115 0.012667 0.004815 0.010678
30 iterations completed
the activations:
−222.586884 −176.979172 −195.530823 −380.166107
−164.0271 −171.654053 −214.053177 −421.249023
−158.297867 −218.833755 −319.384827 −245.097473
−194.550751 −317.505554 −437.527283 −447.651581
the outputs
0.064704 0.10681 0.087355 0.010333
0.122569 0.113061 0.071184 0.006337
0.130157 0.067483 0.021194 0.050156
0.088297 0.021667 0.005218 0.004624
tourcity 0 tour order 1
the tour :
1
0
3
2
distance of tour is : 32

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Output for a Five−City Problem

Please type number of cities, number of iterations
5 40
0.0645 0.069 0.0595 0.0615 0.0825 //input vector—there are 25

neurons in the network
0.074 0.0865 0.056 0.095 0.06
0.0625 0.0685 0.099 0.0645 0.0615
0.056 0.099 0.065 0.093 0.051
0.0675 0.094 0.0595 0.0635 0.0515

10
14
7
6

12
9
18

24
16

32
Distance Matrix
0 10 14 7 6
10 0 12 9 18
14 12 0 24 16
7 9 24 0 32
6 18 16 32 0
Weight Matrix //25x25 matrix of weights. There are 25 neurons

in the network.
−30 −70 −70 −70 −70

−70 −630 −30 −30 −630
−70 −70 −30 −70 −70
−70 −630 −70 −630 −30
−30 −870 −70 −70 −30
−70 −30 −70 −70 −70

−630 −70 −630 −30 −30
−870 −70 −70 −30 −70
−70 −30 −630 −70 −630
−30 −30 −70 −70 −70
−70 −70 −30 −70 −70

−30 −630 −70 −630 −30
−30 −70 −70 −70 −30
−70 −30 −30 −630 −70
−630 −30 −70 −70 −70
−70 −70 −70 −30 −70

−30 −30 −630 −70 −630
−30 −70 −70 −70 −70
−30 −630 −30 −30 −630
−70 −870 −70 −630 −30
−70 −70 −70 −70 −30

−630 −30 −30 −630 −70
−870 −70 −630 −30 −30
−630 −30 −70 −70 −70
−70 −70 −630 −70 −630
−70 −630 −30 −30 −630

−30 −70 −70 −70 −70
−70 −630 −70 −630 −30
−30 −70 −30 −70 −70
−70 −750 −30 −630 −70
−630 −70 −630 −30 −30

−70 −30 −70 −70 −70
−750 −30 −630 −70 −630
−30 −70 −70 −30 −70
−70 −30 −30 −30 −630
−30 −630 −70 −630 −30

−70 −70 −30 −70 −70
−30 −30 −30 −630 −70
−630 −70 −70 −70 −30
−70 −30 −630 −30 −30
−30 −30 −630 −70 −630

−70 −70 −70 −30 −70
−30 −630 −30 −30 −630
−70 −70 −70 −70 −70
−30 −750 −70 −870 −30
−630 −30 −30 −630 −70

−70 −70 −70 −70 −30
−750 −70 −870 −30 −30
−870 −70 −750 −30 −30
−750 −30 −870 −70 −870
−70 −870 −30 −30 −870

−70 −750 −30 −30 −750
−30 −870 −70 −870 −30
−30 −750 −70 −750 −30
−30 −70 −30 −870 −70
−870 −70 −870 −30 −30
−750 −70 −750 −30 −30

−70 −30 −870 −70 −870
−30 −30 −750 −70 −750
−30 −70 −30 −30 −870
−30 −870 −70 −870 −30

−30 −750 −70 −750 −30
−70 −30 −30 −870 −70
−870 −30 −30 −750 −70
−750 −70 −870 −30 −30
−30 −30 −870 −70 −870

−30 −30 −750 −70 −750
−70 −870 −30 −30 −870
−70 −750 −30 −30 −750
−70 −70 −70 −450 −30
−870 −30 −30 −870 −70

−750 −30 −30 −750 −70
−70 −70 −450 −30 −30
−450 −70 −570 −30 −30
−570 −70 −450 −70 −450
−70 −450 −30 −30 −450

−70 −570 −30 −30 −570
−70 −450 −70 −450 −30
−30 −570 −70 −570 −30
−30 −1470 −30 −450 −70
−450 −70 −450 −30 −30

−570 −70 −570 −30 −30
−1470 −30 −450 −70 −450
−30 −30 −570 −70 −570
−30 −30 −30 −30 −450
−30 −450 −70 −450 −30

−30 −570 −70 −570 −30
−30 −30 −30 −450 −70
−450 −30 −30 −570 −70
−570 −30 −450 −30 −30
−30 −30 −450 −70 −450

−30 −30 −570 −70 −570
−30 −450 −30 −30 −450
−70 −570 −30 −30 −570
−70 −1470 −70 −390 −30
−450 −30 −30 −450 −70

−570 −30 −30 −570 −70
−1470 −70 −390 −30 −30
−390 −70 −1110 −30 −30
−1110 −70 −390 −70 −390
−70 −390 −30 −30 −390

−70 −1110 −30 −30 −1110
−70 −390 −70 −390 −30
−30 −1110 −70 −1110 −30
−30 −990 −30 −390 −70
−390 −70 −390 −30 −30

−1110 −70 −1110 −30 −30
−990 −30 −390 −70 −390
−30 −30 −1110 −70 −1110

−30 −30 −30 −30 −390
−30 −390 −70 −390 −30

−30 −1110 −70 −1110 −30
−30 −30 −30 −390 −70
−390 −30 −30 −1110 −70
−1110 −30 −390 −30 −30
−30 −30 −390 −70 −390

−30 −30 −1110 −70 −1110
−30 −390 −30 −30 −390
−70 −1110 −30 −30 −1110
−70 −990 −30 −30 −990
−390 −30 −30 −390 −70

−1110 −30 −30 −1110 −70
−990 −30 −30 −990 −70
−1950 −30 −30 −1950 −70
−70 −70 −70 −70 −30
initial activations
the activations:
−290.894989 −311.190002 −218.365005 −309.344971 −467.774994
−366.299957 −421.254944 −232.399963 −489.249969 −467.399994
−504.375 −552.794983 −798.929871 −496.005005 −424.964935
−374.639984 −654.389832 −336.049988 −612.870056 −405.450012
−544.724976 −751.060059 −418.285034 −545.465027 −500.065063
the outputs
0.029577 0.023333 0.067838 0.023843 0.003636
0.012181 0.006337 0.057932 0.002812 0.003652
0.002346 0.001314 6.859939e−05 0.002594 0.006062
0.011034 0.000389 0.017419 0.000639 0.00765
0.001447 0.000122 0.006565 0.001434 0.002471
40 iterations completed
the activations:
−117.115494 −140.58519 −85.636215 −158.240143 −275.021301
−229.135956 −341.123871 −288.208496 −536.142212 −596.154297
−297.832794 −379.722595 −593.842102 −440.377625 −442.091064
−209.226883 −447.291016 −283.609589 −519.441101 −430.469696
−338.93219 −543.509766 −386.950531 −538.633606 −574.604492
the outputs
0.196963 0.156168 0.263543 0.130235 0.035562
0.060107 0.016407 0.030516 0.001604 0.000781
0.027279 0.010388 0.000803 0.005044 0.004942
0.07511 0.004644 0.032192 0.001959 0.005677
0.016837 0.001468 0.009533 0.001557 0.001012
the tour :
1
2
0
4
3
distance of tour is : 73

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Other Approaches to Solve the Traveling Salesperson Problem
The following describes a few other methods for solving the traveling salesperson problem.
Anzai’s Presentation
Yuichiro Anzai describes the Hopfield network for the traveling salesperson problem in a slightly different
way. For one thing, a global inhibition term is not used. A threshold value is associated with each neuron,
added to the activation, and taken as the average of A1 and A2, using our earlier notation. The energy function
is formulated slightly differently, as follows:
E = A1 £1(£k xik−1)2 + A2 £i(£k xki−1)2 + A4 £k £j`k £i`k dkj xki(xj,i+1 + xj,i−1)
The first term is 0 if the sum of the outputs is 1 in each column. The same is true for the second term with
regard to rows.
The output is calculated using a parameter », here called the reference activation level, as:
xij = (1 + tan tanh(aij/»))/2
The parameters used are A1 = 1/2, A2 = 1/2, A4 = 1/2, ”t = 1, Ä = 1, and » = 1. An attempt is made to solve the
problem for a tour of 10 cities. The solution obtained is not crisp, in the sense that exactly one 1 occurs in
each row and each column, there are values of varying magnitude with one dominating value in each row and
column. The prominent value is considered to be part of the solution.
Kohonen’s Approach for the Traveling Salesperson Problem

Kohonen’s self−organizing maps can be used for the traveling salesperson problem. We summarize the
discussion of this approach described in Eric Davalo’s and Patrick Naim’s work. Each city considered for the
tour is referenced by its x and y coordinates. To each city there corresponds a neuron. The neurons are placed
in a single array, unlike the two−dimensional array used in the Hopfield approach. The first and the last
neurons in the array are considered to be neighbors.
There is a weight vector for each neuron, and it also has two components. The weight vector of a neuron is the
image of the neuron in the map, which is sought to self−organize. There are as many input vectors as there are
cities, and the coordinate pair of a city constitutes an input vector. A neuron with a weight vector closest to the
input vector is selected. The weights of neurons in a neighborhood of the selected neuron are modified, others
are not. A gradually reducing scale factor is also used for the modification of weights.
One neuron is created first, and its weight vector has 0 for its components. Other neurons are created one at a
time, at each iteration of learning. Neurons may also be destroyed. The creation of the neuron and destruction
of the neuron suggest adding a city provisionally to the final list in the tour and dropping a city also
Other Approaches to Solve the Traveling Salesperson Problem 364

provisionally from that list. Thus the possibility of assigning any neuron to two inputs or two cities is
prevented. The same is true about assigning two neurons to the same input.
As the input vectors are presented to the network, if an unselected neuron falls in the neighborhood of the
closest twice, it is created. If a created neuron is not selected in three consecutive iterations for modification of
weights, along with others being modified, it is destroyed.
That a tour of shortest distance results from this network operation is apparent from the fact that the closest
neurons are selected. It is reported that experimental results are very promising. The computation time is
small, and solutions somewhat close to the optimal are obtained, if not the optimal solution itself. As was
before about the traveling salesperson problem, this is an NP−complete problem and near efficient (leading to
suboptimal solutions, but faster) approaches to it should be accepted.
Algorithm for Kohonen’s Approach

A gain parameter » and a scale factor q are used while modifying the weights. A value between 0.02 and 0.2
was tried in previous examples for q. A distance of a neuron from the selected neuron is defined to be an
integer between 0 and n – 1, where n is the number of cities for the tour. This means that these distances are
not necessarily the actual distances between the cities. They could be made representative of the actual
distances in some way. One such attempt is described in the following discussion on C++ implementation.
This distance is denoted by dj for neuron j. A squashing function similar to the Gaussian density function is
also used.
The details of the algorithm are in a paper referenced in Davalo. The steps of the algorithm to the extent given
by Davalo are:
• Find the weight vector for which the distance from the input vector is the smallest
• Modify the weights using
wjnew = wjold + (lnew − wjold)g(»,dj), where g(»,dj) = exp(− dj2/») /
• Reset » as » (1 − q)


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
C++ Implementation of Kohonen’s Approach

Our C++ implementation of this algorithm (described above) is with small modifications. We create but do
not destroy neurons explicitly. That is, we do not count the number of consecutive iterations in which a
neuron is not selected for modification of weights. This is a consequence of our not defining a neighborhood
of a neuron. Our example is for a problem with five neurons, for illustration, and because of the small number
of neurons involved, the entire set is considered a neighborhood of each neuron.
When all but one neuron are created, the remaining neuron is created without any more work with the
algorithm, and it is assigned to the input, which isn’t corresponded yet to a neuron. After creating n – 1
neurons, only one unassigned input should remain.
In our C++ implementation, the distance matrix for the distances between neurons, in our example, is given as
follows, following the stipulation in the algorithm that these values should be integers between 0 and n – 1.
0 1 2 3 4
1 0 1 2 3
d = 2 1 0 1 2
3 2 1 0 1
4 3 2 1 0
We also ran the program by replacing the previous matrix with the following matrix and obtained the same
solution. The actual distances between the cities are about four times the corresponding values in this matrix,
more or less. We have not included the output from this second run of the program.
0 1 3 3 2
1 0 3 2 1
d = 3 3 0 4 2
3 2 4 0 1
2 1 2 1 0
In our implementation, we picked a function similar to the Gaussian density function as the squashing
function. The squashing function used is:
f(d,») = exp( −d2/») / (2 )
Header File for C++ Program for Kohonen’s Approach

Listing 15.3 contains the header file for this program, and listing 15.4 contains the corresponding source file:
Listing 15.3 Header file for C++ program for Kohonen’s approach
//tsp_kohn.h V.Rao, H.Rao

#include<iostream.h>

#include<math.h>
#define MXSIZ 10
#define pi 3.141592654
class city_neuron
{
protected:
double x,y;
int mark,order,count;
double weight[2];
friend class tspnetwork;
public:
city_neuron(){};
void get_neuron(double,double);
};
class tspnetwork
{
protected:
int chosen_city,order[MXSIZ];
double gain,input[MXSIZ][2];
int citycount,index,d[MXSIZ][MXSIZ];
double gain_factor,diffsq[MXSIZ];
city_neuron (cnrn)[MXSIZ];
public:
tspnetwork(int,double,double,double,double*,double*);
void get_input(double*,double*);
void get_d();
void find_tour();
void associate_city();
void modify_weights(int,int);
double wtchange(int,int,double,double);
void print_d();
void print_input();
void print_weights();
void print_tour();
};
Source File Listing

The following is the source file listing for the Kohonen approach to the traveling salesperson problem.
Listing 15.4 Source file for C++ program for Kohonen’s approach
//tsp_kohn.cpp V.Rao, H.Rao

#include “tsp_kohn.h”
void city_neuron::get_neuron(double a,double b)

{
x = a;
y = b;
mark = 0;
count = 0;
weight[0] = 0.0;
weight[1] = 0.0;
};
tspnetwork::tspnetwork(int k,double f,double q,double h,

double *ip0,double *ip1)

{
int i;
gain = h;
gain_factor = f;
citycount = k;
// distances between neurons as integers between 0 and n−1
get_d();
print_d();
cout<<”\n”;
// input vectors
get_input(ip0,ip1);
print_input();
// neurons in the network
for(i=0;i<citycount;++i)
{
order[i] = citycount+1;
diffsq[i] = q;
cnrn[i].get_neuron(ip0[i],ip1[i]);
cnrn[i].order = citycount +1;
}
}
void tspnetwork::associate_city()
{
int i,k,j,u;
double r,s;
for(u=0;u<citycount;u++)
{
//start a new iteration with the input vectors
for(j=0;j<citycount;j++)
{
{
if(cnrn[i].mark==0)
{
k = i;
i =citycount;
}
}
//find the closest neuron
{
r = input[j][0] − cnrn[i].weight[0];
s = input[j][1] − cnrn[i].weight[1];
diffsq[i] = r*r +s*s;
if(diffsq[i]<diffsq[k]) k=i;
}
chosen_city = k;
cnrn[k].count++;
if((cnrn[k].mark<1)&&(cnrn[k].count==2))
{

//associate a neuron with a position

cnrn[k].mark = 1;
cnrn[k].order = u;
order[u] = chosen_city;
index = j;
gain *= gain_factor;
//modify weights
modify_weights(k,index);
print_weights();
j = citycount;
}
}
}
}
void tspnetwork::find_tour()
{
int i;
{
associate_city();
}
//associate the last neuron with remaining position in

// tour
{
if( cnrn[i].mark ==0)
{
cnrn[i].order = citycount−1;
order[citycount−1] = i;
cnrn[i].mark = 1;
}
}
//print out the tour.

//First the neurons in the tour order
//Next cities in the tour
//order with their x,y coordinates
print_tour();
}
void tspnetwork::get_input(double *p,double *q)

{
int i;
{
input[i][0] = p[i];
input[i][1] = q[i];
}
}
//function to compute distances (between 0 and n−1) between

//neurons
void tspnetwork::get_d()
{
int i,j;

{
for(j=0;j<citycount;++j)
{
d[i][j] = (j−i);
if(d[i][j]<0) d[i][j] = d[j][i];
}
}
}
//function to find the change in weight component
double tspnetwork::wtchange(int m,int l,double g,double h)

{
double r;
r = exp(−d[m][l]*d[m][l]/gain);
r *= (g−h)/sqrt(2*pi);
return r;
}
//function to determine new weights
void tspnetwork::modify_weights(int jj,int j)

{
int i;
double t;
double w[2];
{
w[0] = cnrn[i].weight[0];
w[1] = cnrn[i].weight[1];
//determine new first component of weight
t = wtchange(jj,i,input[j][0],w[0]);
w[0] = cnrn[i].weight[0] +t;
cnrn[i].weight[0] = w[0];
//determine new second component of weight

t = wtchange(jj,i,input[j][1],w[1]);
w[1] = cnrn[i].weight[1] +t;
cnrn[i].weight[1] = w[1];
}
}
//different print routines
void tspnetwork::print_d()
{
int i,j;
cout<<”\n”;
for(i=0;i<citycount;i++)
{
cout<<” d: “;
for(j=0;j<citycount;j++)
{
cout<<d[i][j]<<” “;
}
cout<<”\n”;
}
}
void tspnetwork::print_input()

{
int i,j;
{
cout<<”input : “;
for(j=0;j<2;j++)
{
cout<<input [i][j]<<” “;
}
cout<<”\n”;
}
}
void tspnetwork::print_weights()
{
int i,j;
cout<<”\n”;
{
cout<<” weight: “;
for(j=0;j<2;j++)
{
cout<<cnrn[i].weight[j]<<” “;
}
cout<<”\n”;
}
}
void tspnetwork::print_tour()
{
int i,j;
cout<<”\n tour : “;
{
cout<<order[i]<<” —> “;
}
cout<<order[0]<<”\n\n”;
{
j = order[i];
cout<<”(“<<cnrn[j].x<<”, “<<cnrn[j].y<<”) —> “;
}
j= order[0];
cout<<”(“<<cnrn[j].x<<”, “<<cnrn[j].y<<”)\n\n”;
}
void main()
{
int nc= 5;//nc = number of cities
double q= 0.05,h= 1.0,p= 1000.0;
double input2[][5]= {7.0,4.0,14.0,0.0,5.0,3.0,6.0,13.0,12.0,10.0};

tspnetwork tspn2(nc,q,p,h,input2[0],input2[1]);
tspn2.find_tour();



by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Output from a Sample Program Run

The program, as mentioned, is created for the Kohonen approach to the traveling salesperson problem for five
cities. There is no user input from the keyboard. All parameter values are given to the program with
appropriate statements in the function main. A scale factor of 0.05 is given to apply to the gain parameter,
which is given as 1. Initially, the distance of each neuron weight vector from an input vector is set at 1000, to
facilitate finding the closest for the first time. The cities with coordinates (7,3), (4,6), (14,13), (0,12), (5,10)
are specified for input vectors.
The tour found is not the one in natural order, namely 0 ’ 1 ’ 2 ’ 3 ’ 4 ’ 0, with a distance of 43.16. The tour
found has the order 0 ’ 3 ’ 1 ’ 4 ’ 2 ’ 0, which covers a distance of 44.43, which is slightly higher, as shown in
Figure 15.2. The best tour, 0 ’ 2 ’ 4 ’ 3 ’ 1 ’ 0 has a total distance of 38.54.
Figure 15.2 City placement and tour found for TSP.
Table 15.2 gives for the five−city example, the 12 (5!/10) distinct tour distances and corresponding
representative tours. These are not generated by the program, but by enumeration and calculation by hand.
This table is provided here for you to see the different solutions for this five−city example of the traveling
salesperson problem.
Table 15.2 Distances and Representative Tours for Five−City Example

DistanceTourComment
49.050−3−2−1−4−0worst case
47.590−3−1−2−4−0
45.330−2−1−4−3−0
44.860−2−3−1−4−0
44.430−3−1−4−2−0tour given by the program
44.300−2−1−3−4−0
43.290−1−4−2−3−0
43.160−1−2−3−4−0
42.730−1−2−4−3−0
42.260−1−3−2−4−0
40.000−1−4−3−2−0
38.540−2−4−3−1−0optimal tour
There are 12 different distances you can get for tours with these cities by hand calculation, and four of these
are higher and seven are lower than the one you find from this program. The worst case tour (0 [rarr] 3 [rarr] 2

[rarr] 1 [rarr] 4 [rarr] 0) gives a distance of 49.05, and the best, as you saw above, 38.54. The solution from
the program is at about the middle of the best and worst, in terms of total distance traveled.
The output of the program being all computer generated is given below as follows:
d: 0 1 2 3 4
d: 1 0 1 2 3
d: 2 1 0 1 2
d: 3 2 1 0 1
d: 4 3 2 1 0
input : 7 3
input : 4 6
input : 14 13
input : 0 12
input : 5 10
weight: 1.595769 2.393654

weight: 3.289125e−09 4.933688e−09
weight: 2.880126e−35 4.320189e−35
weight: 1.071429e−78 1.607143e−78
weight: 1.693308e−139 2.539961e−139
weight: 1.595769 2.393654

weight: 5.585192 5.18625
weight: 2.880126e−35 4.320189e−35
weight: 1.071429e−78 1.607143e−78
weight: 1.693308e−139 2.539961e−139
weight: 1.595769 2.393654

weight: 5.585192 5.18625
weight: 5.585192 5.18625
weight: 1.071429e−78 1.607143e−78
weight: 1.693308e−139 2.539961e−139
weight: 1.595769 2.393654

weight: 5.585192 5.18625
weight: 5.585192 5.18625
weight: 5.585192 5.18625
weight: 1.693308e−139 2.539961e−139
weight: 1.595769 2.393654

weight: 5.585192 5.18625
weight: 5.585192 5.18625
weight: 5.585192 5.18625
weight: 5.585192 5.18625
tour : 0 –> 3–> 1 –> 4 –> 2–> 0
(7, 3) –> (0, 12) –> (4, 6) –> (5, 10) –> (14, 13) –> (7, 3)
Optimizing a Stock Portfolio

Development of a neural network approach to a stock selection process in securities trading is similar to the
application of neural networks to nonlinear optimization problems. The seminal work of Markowitz in making
a mathematical formulation of an objective function in the context of portfolio selection forms a basis for
such a development. There is risk to be minimized or put a cap on, and there are profits to be maximized.
Investment capital is a limited resource naturally.
Optimizing a Stock Portfolio 374

The objective function is formulated in such a way that the optimal portfolio minimizes the objective
function. There would be a term in the objective function involving the product of each pair of stock prices.
The covariance of that pair of prices is also used in the objective function. A product renders the objective
function a quadratic. There would of course be some linear terms as well, and they represent the individual
stock prices with the stock’s average return as coefficient in each such term. You already get the idea that this
optimization problem falls into the category of quadratic programming problems, which result in real number
values for the variables in the optimal solution. Some other terms would also be included in the objective
function to make sure that the constraints of the problem are satisfied.
A practical consideration is that a real number value for the amount of a stock may be unrealistic, as fractional
numbers of stocks may not be purchased. It makes more sense to ask that the variables be taking 0 or 1 only.
The implication then is that either you buy a stock, in which case you include it in the portfolio, or you do not
buy at all. This is what is usually called a zero−one programming problem. You also identify it as a
combinatorial problem.
You already saw a combinatorial optimization problem in the traveling salesperson problem. The constraints
were incorporated into special terms in the objective function, so that the only function to be computed is the
objective function.
Deeming the objective function as giving the energy of a network in a given state, the simulated annealing
paradigm and the Hopfield network can be used to solve the problem. You then have a neural network in
which each neuron represents a stock, and the size of the layer is determined by the number of stocks in the
pool from which you want to build your stock portfolio. The paradigm suggested here strives to minimize the
energy of the machine. The objective function needs therefore to be stated for minimization to get the best
portfolio possible.
Optimizing a Stock Portfolio 375


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Tabu Neural Network

Tabu search, popularized by Fred Glover with his contributions, is a paradigm that has been used successfully
in many optimization problems. It is a method that can steer a search procedure from a limited domain to an
extended domain, so as to seek a solution that is better than a local minimum or a local maximum.
Tabu search (TS), suggests that an adaptive memory and a responsive exploration need to be part of an
algorithm. Responsive exploration exploits the information derivable from a selected strategy. Such
information may be more substantial, even if the selected strategy is in some sense a bad strategy, than what
you can get even in a good strategy that is based on randomness. It is because there is an opportunity provided
by such information to intelligently modify the strategy. You can get some clues as to how you can modify
the strategy.
When you have a paradigm that incorporates adaptive memory, you see the relevance of associating a neural
network:. a TANN is a Tabu neural network. Tabu search and Kohonen’s self−organizing map have a
common approach in that they work with “neighborhoods.” As a new neighborhood is determined, TS
prohibits some of the earlier solutions, as it classifies them as tabu. Such solutions contain attributes that are
identified as tabu active.
Tabu search, has STM and LTM components as well. The short−term memory is sometimes called
recency−based memory. While this may prove adequate to find good solutions, the inclusion of long−term
memory makes the search method much more potent. It also does not necessitate longer runs of the search
process.
Some of the examples of applications using Tabu search are:
• Training neural nets with the reactive Tabu search

• Tabu Learning: a neural network search method for solving nonconvex optimization problems
• Massively parallel Tabu search for the quadratic assignment problem
• Connection machine implementation of a Tabu search algorithm for the traveling salesman
problem
• A Tabu search procedure for multicommodity location/allocation with balancing requirements
Summary
The traveling salesperson problem is presented in this chapter as an example of nonlinear optimization with
neural networks. Details of formulation are given of the energy function and its evaluation. The approaches to
the solution of the traveling salesperson problem using a Hopfield network and using a Kohonen
self−organizing map are presented. C++ programs are included for both approaches.
The output with the C++ program for the Hopfield network refers to examples of four− and five−city tours.
The output with the C++ program for the Kohonen approach is given for a tour of five cities, for illustration.
Tabu Neural Network 376

The solution obtained is good, if not optimal. The problem with the Hopfield approach lies in the selection of
appropriate values for the parameters. Hopfield’s choices are given for his 10−city tour problem. The same
values for the parameters may not work for the case of a different number of cities. The version of this
approach given by Anzai is also discussed briefly.
Use of neural networks for nonlinear optimization as applied to portfolio selection is also presented in this
chapter. You are introduced to Tabu search and its use in optimization with neural computing.
Tabu Neural Network 377


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Chapter 16
Introduction
Up until now, we have discussed how fuzzy logic could be used in conjunction with neural networks: We
looked at a fuzzifier in Chapter 3 that takes crisp input data and creates fuzzy outputs, which then could be
used as inputs to a neural network. In chapter 9, we used fuzzy logic to create a special type of associative
memory called a FAM (fuzzy associative memory). In this chapter, we focus on applications of fuzzy logic by
itself. This chapter starts with an overview of the different types of application areas for fuzzy logic. We then
present two application domains of fuzzy logic: fuzzy control systems, and fuzzy databases and quantification.
In these sections, we also introduce some more concepts in fuzzy logic theory.
A Fuzzy Universe of Applications
Fuzzy logic is being applied to a wide variety of problems. The most pervasive field of influence is in control
systems, with the rapid acceptance of fuzzy logic controllers (FLCs) for machine and process control. There
are a number of other areas where fuzzy logic is being applied. Here is a brief list adapted from Yan, et al.,
with examples in each area:
• Biological and Medical Sciences Fuzzy logic based diagnosis systems, cancer research, fuzzy
logic based manipulation of prosthetic devices, fuzzy logic based analysis of movement disorders, etc.
• Management and Decision Support Fuzzy logic based factory site selection, fuzzy logic aided
military decision making (sounds scary, but remember that the fuzzy in fuzzy logic applies to the
imprecision in the data and not in the logic), fuzzy logic based decision making for marketing
strategies, etc.
• Economics and Finance Fuzzy modeling of complex marketing systems, fuzzy logic based
trading systems, fuzzy logic based cost−benefit analysis, fuzzy logic based investment evaluation, etc.
• Environmental Science Fuzzy logic based weather prediction, fuzzy logic based water quality
control, etc.
• Engineering and Computer Science Fuzzy database systems, fuzzy logic based prediction of
earthquakes, fuzzy logic based automation of nuclear plant control, fuzzy logic based computer
network design, fuzzy logic based evaluation of architectural design, fuzzy logic control systems, etc.
• Operations Research Fuzzy logic based scheduling and modeling, fuzzy logic based allocation
of resources, etc.
• Pattern Recognition and Classification Fuzzy logic based speech recognition, fuzzy logic based
handwriting recognition, fuzzy logic based facial characteristic analysis, fuzzy logic based military
command analysis, fuzzy image search, etc.
• Psychology Fuzzy logic based analysis of human behavior, criminal investigation and prevention
based on fuzzy logic reasoning, etc.
Chapter 16 Applications of Fuzzy Logic 378

• Reliability and Quality Control Fuzzy logic based failure diagnosis, production line monitoring
and inspection, etc.
We will now move to one of the two application domains that we will discuss in depth, Fuzzy Databases.
Later in the chapter, we examine the second application domain, Fuzzy Control Systems.
Section I: A Look at Fuzzy Databases and Quantification

In this section, we want to look at some ways in which fuzzy logic may be applied to databases and operations
with databases. Standard databases have crisp data sets, and you create unambiguous relations over the data
sets. You also make queries that are specific and that do not have any ambiguity. Introducing ambiguity in one
or more of these aspects of standard databases leads to ideas of how fuzzy logic can be applied to databases.
Such application of fuzzy logic could mean that you get databases that are easier to query and easier to
interface to. A fuzzy search, where search criteria are not precisely bounded, may be more appropriate than a
crisp search. You can recall any number of occasions when you tend to make ambiguous queries, since you
are not certain of what you need. You also tend to make ambiguous queries when a “ball park” value is
sufficient for your purposes.
In this section, you will also learn some more concepts in fuzzy logic. You will see these concepts introduced
where they arise in the discussion of ideas relating to fuzzy databases. You may at times see somewhat of a
digression in the middle of the fuzzy database discussion to fuzzy logic topics. You may skip to the area
where the fuzzy database discussion is resumed and refer back to the skipped areas whenever you feel the
need to get clarification of a concept.
We will start with an example of a standard database, relations and queries. We then point out some of the
ways in which fuzziness can be introduced.
Section I: A Look at Fuzzy Databases and Quantification 379


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Databases and Queries
Imagine that you are interested in the travel business. You may be trying to design special tours in different
countries with your own team of tour guides, etc. , and you want to identify suitable persons for these
positions. Initially, let us say, you are interested in their own experiences in traveling, and the knowledge they
possess, in terms of geography, customs, language, and special occasions, etc. The information you want to
keep in your database may be something like, who the person is, the person’s citizenship, to where the person
traveled, when such travel occurred, the length of stay at that destination, the person’s languages, the
languages the person understands, the number of trips the person made to each place of travel, etc. Let us use
some abbreviations:
cov—country visited
lov—length of visit (days)
nov—number of visits including previous visits
ctz—citizenship
yov—year of visit
lps—language (other than mother tongue) with proficiency to speak
lpu—language with only proficiency to understand
hs—history was studied (1—yes, 0—no)
Typical entries may appear as noted in Table 16.1.
Table 16.1 Example Database

Nameagectzcovlovnovyovlpslpuhs
John Smith35U.S.India411994Hindi1
John Smith35U.S.Italy721991Italian1
John Smith35U.S.Japan3119930
When a query is made to list persons that visited India or Japan after 1992 for 3 or more days, John Smith’s
two entries will be included. The conditions stated for this query are straightforward, with lov [ge] 3 and yov
> 1992 and (cov = India or cov = Japan).
Relations in Databases
A relation from this database may be the set of quintuples, (name, age, cov, lov, yov). Another may be the set
of triples, (name, ctz, lps). The quintuple (John Smith, 35, India, 4, 1994) belongs to the former relation, and
the triple (John Smith, U.S., Italian) belongs to the latter. You can define other relations, as well.
Databases and Queries 380

Fuzzy Scenarios
Now the query part may be made fuzzy by asking to list young persons who recently visited Japan or India for
a few days. John Smith’s entries may or may not be included this time since it is not clear if John Smith is
considered young, or whether 1993 is considered recent, or if 3 days would qualify as a few days for the
query. This modification of the query illustrates one of three scenarios in which fuzziness can be introduced
into databases and their use.
This is the case where the database and relations are standard, but the queries may be fuzzy. The other cases
are: one where the database is fuzzy, but the queries are standard with no ambiguity; and one where you have
both a fuzzy database and some fuzzy queries.
Fuzzy Sets Revisited
We will illustrate the concept of fuzziness in the case where the database and the queries have fuzziness in
them. Our discussion is guided by the reference Terano, Asai, and Sugeno. First, let us review and recast the
concept of a fuzzy set in a slightly different notation.
If a, b, c, and d are in the set A with 0.9, 0.4, 0.5, 0, respectively, as degrees of membership, and in B with
0.9, 0.6, 0.3, 0.8, respectively, we give these fuzzy sets A and B as A = { 0.9/a, 0.4/b, 0.5/c} and B = {0.9/a,
0.6/b, 0.3/c, 0.8/d}. Now A[cup]B = {0.9/a, 0.6/b, 0.5/c, 0.8/d} since you take the larger of the degrees of
membership in A and B for each element. Also, A[cap]B = {0.9/a, 0.4/b, 0.3/c} since you now take the
smaller of the degrees of membership in A and B for each element. Since d has 0 as degree of membership in
A (it is therefore not listed in A), it is not listed in A[cap]B.
Let us impart fuzzy values (FV) to each of the attributes, age, lov, nov, yov, and hs by defining the sets in
Table 16.2.
Table 16.2 Fuzzy Values for Example Sets

Fuzzy ValueSet
FV(age){ very young, young, somewhat old, old }
FV(nov){ never, rarely, quite a few, often, very often }
FV(lov){ barely few days, few days, quite a few days, many days }
FV(yov){distant past, recent past, recent }
FV(hs){ barely, adequately, quite a bit, extensively }
The attributes of name, citizenship, country of visit are clearly not candidates for having fuzzy values. The
attributes of lps, and lpu, which stand for language in which speaking proficiency and language in which
understanding ability exist, can be coupled into another attribute called flp (foreign language proficiency) with
fuzzy values. We could have introduced in the original list an attribute called lpr ( language with proficiency
to read) along with lps and lpu. As you can see, these three can be taken together into the fuzzy−valued
attribute of foreign language proficiency. We give below the fuzzy values of flp.
FV(flp) = {not proficient, barely proficient, adequate,

proficient, very proficient }
Note that each fuzzy value of each attribute gives rise to a fuzzy set, which depends on the elements you
consider for the set and their degrees of membership.
Fuzzy Scenarios 381

Fuzzy Scenarios 382


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Now let us determine the fuzzy sets that have John Smith as an element, besides possibly others. We need the
values of degrees of membership for John Smith (actually his attribute values) in various fuzzy sets. Let us
pick them as follows:
——————————————————————————————————————————————————————————————————————————
Age: mvery young(35) = 0 (degree of membership of 35 in very young is 0.
We will employ this notation from now on).
myoung(35) = 0.75
msomewhat old(35) = 0.3
mold(35) = 0
——————————————————————————————————————————————————————————————————————————
Assume that similar values are assigned to the degrees of membership of values of John Smith’s attributes in
other fuzzy sets. Just as John Smith’s age does not belong in the fuzzy sets young and old, some of his other
attribute values do not belong in some of the other fuzzy sets. The following is a list of fuzzy sets in which
John Smith appears:
age_young = {0.75/35, ...}

age_somewhat old = {0.3/35, ... }
A similar statement attempted for the number of visits may prompt you to list nov_rarely = {0.7/1, 0.2/2},
and nov_quite a few = {0.3/2, .6/3, ...}. But you readily realize that the number of visits by itself does not
mean much unless it is referenced with the country of visit. A person may visit one country very often, but
another only rarely. This suggests the notion of a fuzzy relation, which is also a fuzzy set.
NOTE: What follows is an explanation of relations and discussion of fuzzy relations. If

you want to skip this part for now, you may go to the “Fuzzy Queries” section a few pages
later in this chapter.
Fuzzy Relations
A standard relation from set A to set B is a subset of the Cartesian product of A and B, written as A×B. The
elements of A×B are ordered pairs (a, b) where a is an element of A and b is an element of B. For example, the
ordered pair (Joe, Paul) is an element of the Cartesian product of the set of fathers, which includes Joe and the
set of sons which includes Paul. Or, you can consider it as an element of the Cartesian product of the set of
men with itself. In this case, the ordered pair (Joe, Paul) is in the subset which contains (a, b, if a is the father
of b. This subset is a relation on the set of men. You can call this relation “father.”
A fuzzy relation is similar to a standard relation, except that the resulting sets are fuzzy sets. An example of
such a relation is ‘much_more_educated’. This fuzzy set may look something like,
much_more_educated = { ..., 0.2/(Jeff, Steve), 0.7/(Jeff, Mike), ... }
Fuzzy Relations 383

Matrix Representation of a Fuzzy Relation
A fuzzy relation can be given as a matrix also when the underlying sets, call them domains, are finite. For
example, let the set of men be S = { Jeff, Steve, Mike }, and let us use the same relation,
much_more_educated. For each element of the Cartesian product S×S, we need the degree of membership in
this relation. We already have two such values, mmuch_more_educated(Jeff, Steve) = 0.2, and
mmuch_more_educated(Jeff, Mike) = 0.7. What degree of membership in the set should we assign for the pair (Jeff,
Jeff)? It seems reasonable to assign a 0. We will assign a 0 whenever the two members of the ordered pair are
the same. Our relation much_more_educated is given by a matrix that may look like the following:
0/(Jeff, Jeff) 0.2/(Jeff, Steve) 0.7/(Jeff, Mike)

much_more_educated = 0.4/(Steve, Jeff) 0/(Steve, Steve) 0.3/(Steve, Mike)
0.1/(Mike, Jeff) 0.6/(Mike, Steve) 0/(Mike, Mike)
NOTE: Note that the first row corresponds to ordered pairs with Jeff as the first member,
second column corresponds to those with Steve as the second member, and so on. The main
diagonal has ordered pairs where the first and second members are the same.
Matrix Representation of a Fuzzy Relation 384


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Properties of Fuzzy Relations
A relation on a set, that is a subset of a Cartesian product of some set with itself, may have some interesting
properties. It may be reflexive. For this you need to have 1 for the degree of membership of each main
diagonal entry. Our example here is evidently not reflexive.
A relation may be symmetric. For this you need the degrees of membership of each pair of entries
symmetrically situated to the main diagonal to be the same value. For example (Jeff, Mike) and (Mike, Jeff)
should have the same degree of membership. Here they do not, so our example of a relation is not symmetric.
A relation may be antisymmetric. This requires that if a is different from b and the degree of membership of
the ordered pair (a, b) is not 0, then its mirror image, the ordered pair (b, a), should have 0 for degree of
membership. In our example, both (Steve, Mike) and (Mike, Steve) have positive values for degree of
membership; therefore, the relation much_more_educated over the set {Jeff, Steve, Mike} is not
antisymmetric also.
A relation may be transitive. For transitivity of a relation, you need the following condition, illustrated with
our set {Jeff, Steve, Mike}. For brevity, let us use r in place of much_more_educated, the name of the
relation:
min (mr(Jeff, Steve) , mr(Steve, Mike) )[le]mr(Jeff, Mike)

min (mr(Jeff, Mike) , mr(Mike, Steve) )[le]mr(Jeff, Steve)
min (mr(Steve, Jeff) , mr(Jeff, Mike) )[le]mr(Steve, Mike)
min (mr(Steve, Mike) , mr(Mike, Jeff) )[le]mr(Steve, Jeff)
min (mr(Mike, Jeff) , mr(Jeff, Steve) )[le]mr(Mike, Steve)
min (mr(Mike, Steve) , mr(Steve, Jeff) )[le]mr(Mike, Jeff)
In the above listings, the ordered pairs on the left−hand side of an occurrence of [le] are such that the second
member of the first ordered pair matches the first member of the second ordered pair, and also the right−hand
side ordered pair is made up of the two nonmatching elements, in the same order.
In our example,
min (mr(Jeff, Steve) , mr(Steve, Mike) ) = min (0.2, 0.3) = 0.2

mr(Jeff, Mike) = 0.7 > 0.2
For this instance, the required condition is met. But in the following:
min (mr(Jeff, Mike), mr(Mike, Steve) ) = min (0.7, 0.6) = 0.6

mr(Jeff, Steve) = 0.2 < 0.6
The required condition is violated, so the relation much_more_educated is not transitive.
Properties of Fuzzy Relations 385

NOTE: If a condition defining a property of a relation is not met even in one instance, the
relation does not possess that property. Therefore, the relation in our example is not reflexive,
not symmetric, not even antisymmetric, and not transitive.
If you think about it, it should be clear that when a relation on a set of more than one element is symmetric, it
cannot be antisymmetric also, and vice versa. But a relation can be both not symmetric and not antisymmetric
at the same time, as in our example.
An example of reflexive, symmetric, and transitive relation is given by the following matrix:
1 0.4 0.8
0.4 1 0.4
0.8 0.4 1
Similarity Relations
A reflexive, symmetric, and transitive fuzzy relation is said to be a fuzzy equivalence relation. Such a relation
is also called a similarity relation. When you have a similarity relation s, you can define the similarity class of
an element x of the domain as the fuzzy set in which the degree of membership of y in the domain is ms(x, y).
The similarity class of x with the relation s can be denoted by [x]s.
Resemblance Relations
Do you think similarity and resemblance are one and the same? If x is similar to y, does it mean that x
resembles y? Or does the answer depend on what sense is used to talk of similarity or of resemblance? In
everyday jargon, Bill may be similar to George in the sense of holding high office, but does Bill resemble
George in financial terms? Does this prompt us to look at a ‘resemblance relation’ and distinguish it from the
‘similarity relation’? Of course.
Recall that a fuzzy relation that is reflexive, symmetric, and also transitive is called similarity relation. It helps
you to create similarity classes. If the relation lacks any one of the three properties, it is not a similarity
relation. But if it is only not transitive, meaning it is both reflexive and symmetric, it is still not a similarity
relation, but it is a resemblance relation. An example of a resemblance relation, call it t, is given by the
following matrix.
Let the domain have elements a, b, and c:
1 0.4 0.8
t = 0.4 1 0.5
0.8 0.5 1
This fuzzy relation is clearly reflexive, and symmetric, but it is not transitive. For example:
min (mt(a, c) , mt(c, b) ) = min (0.8, 0.5) = 0.5 ,
but the following:
mt(a, b) = 0.4 < 0.5 ,
is a violation of the condition for transitivity. Therefore, t is not a similarity relation, but it certainly is a
resemblance relation.
Similarity Relations 386

Fuzzy Partial Order
One last definition is that of a fuzzy partial order. A fuzzy relation that is reflexive, antisymmetric, and
transitive is a fuzzy partial order. It differs from a similarity relation by requiring antisymmetry instead of
symmetry. In the context of crisp sets, an equivalence relation that helps to generate equivalence classes is
also a reflexive, symmetric, and transitive relation. But those equivalence classes are disjoint, unlike similarity
classes with fuzzy relations. With crisp sets, you can define a partial order, and it serves as a basis for making
comparison of elements in the domain with one another.
Fuzzy Partial Order 387


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Fuzzy Queries
At this point, our digression from the discussion of fuzzy data bases is finished. Let us now recall, for
immediate reference, the entries in the definitions we listed earlier in Table 16.3:
Table 16.3 Fuzzy Values for Example Sets

Fuzzy ValueSet
FV(age){ very young, young, somewhat old, old }
FV(nov){ never, rarely, quite a few, often, very often }
FV(lov){ barely few days, few days, quite a few days, many days }
FV(yov){distant past, recent past, recent }
FV(hs){ barely, adequately, quite a bit, extensively }
FV(flp){not proficient, barely proficient, adequate, proficient, very proficient }
If you refer to Tables 16.1 and 16.3, you will see that a fuzzy query defines a fuzzy set. For example, suppose
you ask for a list of young people with enough knowledge of Italy, and who know Italian almost like a native.
You want the intersection of the fuzzy sets, age_young, hs_adequately, and flp_very proficient. John Smith
is in the fuzzy set age_young with a 0.75 degree of membership, as noted before. Suppose he is in
hs_adequately with a 0.2 degree of membership, and in flp_very proficient with a 0.9 degree of
membership. We then put him in the fuzzy set for this query. The degree of membership for John Smith is the
smallest value among 0.75, 0.2, and 0.9, since we are taking the intersection of fuzzy sets. So 0.2/John Smith
will be part of the enumeration of the query fuzzy set.
Note that you can use unknown in place of the value of an attribute for an item in the database. If you do not
know John Smith’s age, you can enter unknown in that field of John Smith’s record. You may say his age is
unknown, even though you have a rough idea that he is about 35. You would then be able to assign some
reasonable degrees of membership of John Smith in fuzzy sets like age_young or age_somewhat old.
Extending Database Models
One way of extending a model into a fuzzy model, as far as databases are concerned, is to make use of
similarity relations and to extend the operations with them as Buckles and Perry do, such as PROJECT. First
there are the domains, for the database. In our example relating to travel and tour guides above, the domains
are:
D1 = { John Smith, ... }, the set of persons included in the database,

D2 = {India, Italy, Japan, ... }, the set of countries of visit,
D3 = {Hindi, Italian, Japanese, ... }, the set of foreign languages,
D4 = {U.S., ... }, the set of countries of citizenship,
Fuzzy Queries 388

D5 = set of ages,
D6 = set of years,
D7 = set of number of visits,
D8 = set of length of visits.
Note that the enumerated sets are shown with ‘...’ in the sets, to indicate that there may be more entries listed,
but we are not giving a complete list. In practice, you will make a complete enumeration unlike in our
example. The domains D5, D6, D7, and D8 can also be given with all their elements listed, since in practice,
these sets are finite. Conceptually though, they are infinite sets; for example, D6 is the set of positive integers.
Next, you have the similarity relations that you define. And then there are the operations.
Example
Consider a standard database given below. It can be called also the relation R1, from set D1 = {Georgette,
Darrell, Ernie , Grace } to set D2 = {Spanish, Italian, French, Japanese, Chinese, Russian} as shown in Table
16.4:
Table 16.4 Example for Relation R1, with Domains D1 and D2

D1D2
GeorgetteSpanish
GeorgetteItalian
DarrellFrench
DarrellSpanish
DarrellJapanese
ErnieSpanish
ErnieRussian
GraceChinese
GraceFrench
Let’s say a fuzzy database has D1 and D2 as domains. Suppose you have a similarity relation S1 on domain
D1 given by the matrix:
1 0.4 0.8 0.7

0.4 1 0.5 0.3
0.8 0.5 1 0.4
0.7 0.3 0.4 1
Recall that the entries in this matrix represent the degrees of membership of ordered pairs in the relation S1.
For example, the first row third column element, which is 0.8, in the matrix refers to the degree of
membership of the pair (Georgette, Ernie), since the rows and columns refer to Georgette, Darrell, Ernie and
Grace in that order.
Example 389

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
The result of the operation:
PROJECT (R1 over D1) with LEVEL(D1) = 0.6
is the relation R2 given in Table 16.5.
Table 16.5 Relation R2, which is the Result of PROJECT Operation

D1
{Georgette, Ernie}
{Georgette, Grace}
Darrell
This projection operation with a condition on the level works as follows. First, the column for D1 is to be
picked, of the two columns that R1 shows. Repetitions are removed. The condition says that if two elements
of D1 have a similarity of 0.6 or higher, they should be treated as the same. Even though similarity levels of
pairs (Georgette, Ernie) and (Georgette, Grace) are both greater than 0.6, the pair (Ernie, Grace) has similarity
of 0.4 only so that we do not treat the three as the same.
This type of table is not constructed in the standard database model, so this is part of an extension of the
standard model.
Suppose you recast the information in the relation R1 and call it R3, shown in Table 16.6.
Table 16.6 Example for Relation R3, with Domains D1 and D2

D1D2
Georgette{Spanish, Italian}
Darrell{French, Spanish, Japanese}
Grace{Chinese, French}
Ernie{Russian, Spanish}
This kind of a table also is not found in standard databases (where there are groups with more than one
element used in the relation), and is an example of an extended model.
Possibility Distributions
As an alternative to using similarity relations for introducing fuzziness into a database model, you can,
following Umano, et al., use a possibility distribution−relational model. The possibility distributions
represent the fuzzy values of attributes in the data. An example of a possibility distribution is the fuzzy set you
saw before, nov_rarely = {0.7/1, 0.2/2}, where nov_rarely stands for number of visits considered to be
“rarely.”
Possibility Distributions 390

Example
An example of a database on the lines of this model is shown in Table 16.7:
Table 16.7 Example of Possibility Distribution Relational Model

NameNumber of Visits Outside the U.S.CitizenshipName of Companion on Latest Visit
Peter3U.S.Barbara
Roberto{10, 11}pSpainAnne
Andre2unknownCarol
Raj14{India, U.S.}pUma
AlanunknownU.S.undefined
JamesmanyU.K.null
A standard database cannot look like this. Entries like many and {10, 11}p clearly suggest fuzziness. The
entry {10, 11}p, is a possibility distribution, suggesting that the number of visits outside the United States
made by Roberto is either 10 or 11. Similarly, Raj’s citizenship is India or United States, but not dual
citizenship in both. Andre’s citizenship and Alan’s number of visits outside the United States are not known,
and they can have any values. The possibilities cannot be narrowed down as in the case of Raj’s citizenship
and Roberto’s frequency of visits outside the United States The entry undefined is used for Alan because he
always traveled alone, he never took a companion.
James’ number of visits is fuzzy. He traveled many times. A fuzzy set for many will provide the possibility
distribution. It can be defined, for example, as:
many = {0.2/6, 0.5/7, 0.8/8, 1/9, 1/10, ...}
The name of the companion on James’ latest visit outside the United States is entered as null because we do
not know on the one hand whether he never took a companion, in which case we could have used undefined
as in Alan’s case, and on the other whom he took as a companion if he did take one, in which case we could
have used unknown. Simply put, we use null when we do not know enough to use either unknown or
undefined.
Possibility Distributions 391


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Queries
Let us turn our attention now to how queries are answered with this type of a database model. Suppose you
want a list of U.S. citizens in your database. Peter and Alan clearly satisfy this condition on citizenship. Andre
and Raj can only be said to possibly satisfy this condition. But Roberto and James clearly do not satisfy the
given condition. This query itself is crisp and not fuzzy (either you belong to the list of U.S. citizens or you
don’t). The answer therefore should be a crisp set, meaning that unless the degree of membership is 1, you
will not list an element. So you get the set containing Peter and Alan only.
A second query could be for people who made more than a few visits outside the United States Here the query
is fuzzy. James with many visits outside United States seems to clearly satisfy the given condition. It is
perhaps reasonable to assume that each element in the fuzzy set for many appears in the fuzzy set more than
a few with a degree of membership 1. Andre’s 2 may be a number that merits 0 degree of membership in
more than a few. The other numbers in the database are such that they possibly satisfy the given condition to
different degrees. You can see that the answer to this fuzzy query is a fuzzy set. Now, we switch gears a little,
to talk more on fuzzy theory. This will help with material to follow.
Fuzzy Events, Means and Variances

Let us introduce you to fuzzy events, fuzzy means, and fuzzy variances. These concepts are basic to make a
study of fuzzy quantification theories. You will see how a variable’s probability distribution and its fuzzy set
are used together. We will use an example to show how fuzzy means and fuzzy variances are calculated.
Example: XYZ Company Takeover Price
Suppose you are a shareholder of company XYZ and that you read in the papers that its takeover is a prospect.
Currently the company shares are selling at $40 a share. You read about a hostile takeover by a group
prepared to pay $100 a share for XYZ. Another company whose business is related to XYZ’s business and has
been on friendly terms with XYZ is offering $85 a share. The employees of XYZ are concerned about their
job security and have organized themselves in preparing to buy the company collectively for $60 a share. The
buyout price of the company shares is a variable with these three possible values, viz., 100, 85, and 60. The
board of directors of XYZ have to make the ultimate decision regarding whom they would sell the company.
The probabilities are 0.3, 0.5, and 0.2 respectively, that the board decides to sell at $100 to the first group, to
sell at $85 to the second, and to let the employees buy out at $60.
Thus, you get the probability distribution of the takeover price to be as follows:
price 100 85 60
probability 0.3 0.5 0.2
Queries 392
From standard probability theory, this distribution gives a mean(or expected price) of:
100 x 0.3 + 85 x 0.5 + 60 x 0.2 = 84.5
and a variance of :
(100−84.5)2 x 0.3 +(85−84.5)2 x 0.5 + (60−84.5)2 x 0.2 = 124.825
Suppose now that a security analyst specializing in takeover situations feels that the board hates a hostile
takeover but to some extent they cannot resist the price being offered. The analyst also thinks that the board
likes to keep some control over the company, which is possible if they sell the company to their friendly
associate company. The analyst recognizes that the Board is sensitive to the fears and apprehensions of their
loyal employees with whom they built a healthy relationship over the years, and are going to consider the
offer from the employees.
The analyst’s feelings are reflected in the following fuzzy set:
{0.7/100, 1/85, 0.5/60}
You recall that this notation says, the degree of membership of 100 in this set is 0.7, that of 85 is 1, and the
degree of membership is 0.5 for the value 60.
The fuzzy set obtained in this manner is also called a fuzzy event. A different analyst may define a different
fuzzy event with the takeover price values. You, as a shareholder, may have your own intuition that suggests a
different fuzzy event. Let us stay with the previous fuzzy set that we got from a security analyst, and give it
the name A.
Probability of a Fuzzy Event
At this point, we can calculate the probability for the fuzzy event A by using the takeover prices as the basis to
correspond the probabilities in the probability distribution and the degrees of membership in A. In other
words, the degrees of membership, 0.7, 1, and 0.5 are treated as having the probabilities 0.3, 0.5, and 0.2,
respectively. But we want the probability of the fuzzy event A, which we calculate as the expected degree of
membership under the probability distribution we are using.
Our calculation gives the following:
0.7 x 0.3 + 1 x 0.5 + 0.5 x 0.2 = 0.21 + 0.5 + 0.1 = 0.81
Probability of a Fuzzy Event 393


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Fuzzy Mean and Fuzzy Variance
Our next step is to calculate the fuzzy mean of the takeover price. Let us call it A_fuzzy_mean, to make
reference to the fuzzy event used.
The calculation is as follows:
A_fuzzy_mean = (1/0.81) x (100 x 0.7 x 0.3 + 85 x 1 x 0.5 + 60 x 0.5 x

0.2) = 85.8
To get the fuzzy variance of the takeover price, you need to use values like (100−85.8)2 , which is the square
of the deviation of the takeover price 100 from the fuzzy mean. A simpler way to calculate, which is
mathematically equivalent, is to first take the fuzzy expected value of the square of the takeover price and
then to subtract the square of the fuzzy mean. To make it easier, let us use p as the variable that represents the
takeover price.
The calculations are as below:
A_fuzzy_expected p2 = 1/(0.81) x (1002x 0.7 x 0.3 + 852x 1 x 0.5 + 602x

0.5 x 0.2) = 7496.91
A_fuzzy_variance = 7496.91 − 85.82 = 7496.91 − 7361.64 = 135.27
Fuzzy logic is thus introduced into the realm of probability concepts such as events, and statistical concepts
such as mean and variance. Further, you can talk of fuzzy conditional expectations and fuzzy posterior
probabilities, etc. enabling you to use fuzziness in Bayesian concepts, regression analysis, and so on. You will
then be delving into the field of fuzzy quantification theories. In what follows, we continue our discussion
with fuzzy conditional expectations.
Conditional Probability of a Fuzzy Event
Suppose you, as a shareholder of the XYZ company in the previous example come up with the fuzzy set, we
will call the fuzzy event:
B = {0.8/100, 0.4/85, 0.7/60}
The probability for your fuzzy event is as follows:
0.8 x 0.3 + 0.4 x 0.5 + 0.7 x 0.2 = 0.58
B_fuzzy_mean and B_fuzzy_variance of the takeover price of XYZ stock work out as 85.17 and 244.35,
respectively. But you want to see how these values change if at all when you take A, the analyst’s fuzzy event,
as a given. You are then asking to determine the conditional probability of your fuzzy event, and your
conditional fuzzy mean and fuzzy variance as well.
Fuzzy Mean and Fuzzy Variance 394

The conditional probability of your fuzzy event is calculated as follows:
(1/0.81) x (0.8 x 0.7 x 0.3 + 0.4 x 1 x 0.5 + 0.7 x 0.5 x 0.2) = 0.54
This value is smaller than the probability you got before when you did not take the analyst’s fuzzy event as
given. The a priori probability of fuzzy event B is 0.58, while the a posteriori probability of fuzzy event B
given the fuzzy event A is 0.54.
Conditional Fuzzy Mean and Fuzzy Variance
The conditional B_fuzzy_mean of the takeover price with fuzzy event A as given works out as:
(1/0.54) x (100 x 0.8 x 0.7 x 0.3 + 85 x 0.4 x 1 x 0.5 + 60 x 0.7 x

0.5 x 0.2) = 70.37
and the conditional B_fuzzy_variance of the takeover price with fuzzy event A, as given, amounts to
1301.76, which is over five times as large as when you did not take the analyst’s fuzzy event as given.
Linear Regression a la Possibilities

When you see the definitions of fuzzy means and fuzzy variances, you may think that regression analysis can
also be dealt with in the realm of fuzzy logic. In this section we discuss what approach is being taken in this
regard.
First, recall what regression analysis usually means. You have a set of x− values and a corresponding set of y
values, constituting a number of sample observations on variables X and Y. In determining a linear regression
of Y on X, you are taking Y as the dependent variable, and X as the independent variable, and the linear
regression of Y on X is a linear equation expressing Y in terms of X. This equation gives you the line ‘closest’
to the sample points (the scatter diagram) in some sense. You determine the coefficients in the equation as
those values that minimize the sum of squares of deviations of the actual y values from the y values from the
line. Once the coefficients are determined, you can use the equation to estimate the value of Y for any given
value of X. People use regression equations for forecasting.
Sometimes you want to consider more than one independent variable, because you feel that there are more
than one variable which collectively can explain the variations in the value of the dependent variable. This is
your multiple regression model. Choosing your independent variables is where you show your modeling
expertise when you want to explain what happens to Y, as X varies.
In any case, you realize that it is an optimization problem as well, since the minimization of the sum of
squares of deviations is involved. Calculus is used to do this for Linear Regression. Use of calculus methods
requires certain continuity properties. When such properties are not present, then some other method has to be
used for the optimization problem.
The problem can be formulated as a linear programming problem, and techniques for solving linear
programming problems can be used. You take this route for solving a linear regression problem with fuzzy
logic.
In a previous section, you learned about possibility distributions. The linear regression problem with fuzzy
logic is referred to as a linear possibility regression problem. The model, following the description of it by
Tarano, Asai, and Sugeno, depends upon a reference function L, and fuzzy numbers in the form of ordered
pairs (a, b). We will present fuzzy numbers in the next section and then return to continue our discussion of
Conditional Fuzzy Mean and Fuzzy Variance 395

the linear possibility regression model.
Conditional Fuzzy Mean and Fuzzy Variance 396


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Fuzzy Numbers
A fuzzy number is an ordered pair of numbers (a, b) with respect to a reference function L, which gives you
the membership function. Here b has to be a positive number.
Here are the properties of L, the reference function:
1. It is a function of one variable and is symmetric about 0. That is, L(x) = L(−x).
2. It has a value of 1 at x = 0. In other words, L(0) = 1.
3. It is generally decreasing, when x is 0 or a positive number, meaning that its value drops as the
value of its argument is increased. For example, L(2) < L(1). Thus L(x) has its values less than 1 for
positive values of x. It does not make sense to have negative numbers as values of L, and so you
ignore the values of x that cause such values for L.
4. The maximum value for the function is 1, at x = 0. It has a sort of a bell−shaped curve.
If A is a fuzzy number with the ordered pair (a, b) with b > 0, and if the reference function is L, you do the
following:
You write the fuzzy number as A = (a, b)L. You get the membership of any element x, by taking the quantity
(x − a) / b (which reminds you of how you get a z−score), and evaluate the reference function L at this
argument. That is:
mA(x) = L( (x−a) / b)
Examples of a reference function L are:
Example 1: L(x) = max ( 0, 1− x2)
You obtain the following shown in Table 16.8.
Table 16.8 Reference Function L

xL(x)
−20
−10
−0.50.75
01
0.50.75
10
20
Fuzzy Numbers 397

This function is not strictly decreasing though. It remains a constant at 0 for x > 1.
Example 2: L(x) = 1/ (1 + x2 )
You get the values shown in Table 16.9.
Table 16.9 Reference Function L

xL(x)
−70.02
−20.2
−10.5
−0.50.8
01
0.50.8
10.5
20.2
70.02
Let us now determine the membership of 3 in the fuzzy number A = (4, 10)L, where L is the function in the
second example above, viz., 1/ (1 + x2).
First, you get (x− 4) / 10 = (3 − 4) / 10 = − 0.1. Use this as the argument of the reference function L. 1/ (1 + (−
0.1)2 ) gives 0.99. This is expressed as follows:
mA(3) = L( (3−4) / 10) = 1/ (1 + (− 0.1)2 ) = 0.99
You can verify the values, mA(0) = 0.862, and mA(10) = 0.735.
Triangular Fuzzy Number
With the right choice of a reference function, you can get a symmetrical fuzzy number A, such that when you
plot the membership function in A, you get a triangle containing the pairs (x, mA(x)), with mA(x) > 0. An
example is A = (5, 8)L, where L = max(1 − |x|, 0).
The numbers x that have positive values for mA(x) are in the interval ( −3, 13 ). Also, mA( −3 ) = 0, and
mA(13) = 0. However, mA(x) has its maximum value at x = 5. Now, if x is less than −3 or greater than 13, the
value of L is zero, and you do not consider such a number for membership in A. So all the elements for which
membership in A is nonzero are in the triangle.
This triangular fuzzy number is shown in Figure 16.1. The height of the triangle is 1, and the width is 16,
twice the number 8, midpoint of the base is at 5. The pair of numbers 5 and 8 are the ones defining the
symmetrical fuzzy number A. The vertical axis gives the membership, so the range for this is from 0 to 1.
Figure 16.1 Triangular membership function.
Triangular Fuzzy Number 398

Linear Possibility Regression Model
Assume that you have (n + 1)−tuples of values of x1, ... xn, and y. That is, for each i, i ranging from 1 to k,
you have (x1, ... , xn, y), which are k sample observations on X1, ... , Xn, and Y. The linear possibility
regression model is formulated differently depending upon whether the data collected is crisp or fuzzy.
Let us give such a model below, for the case with crisp data. Then the fuzziness lies in the coefficients in the
model. You use symmetrical fuzzy numbers, Aj = (aj, bj)L. The linear possibility regression model is
formulated as:
Yj = A0 + A1Xj1 + ... + AnXjn
The value of Y from the model is a fuzzy number, since it is a function of the fuzzy coefficients. The fuzzy
coefficients Ajare chosen as those that minimize the width (the base of the triangle) of the fuzzy number Yj.
But Aj is also determined by how big the membership of observed yj is to be, in the fuzzy number Yj . This
last observation provides a constraint for the linear programming problem which needs to be solved to find the
linear possibility regression. You select a value d, and ask that mY(y) [ge] d.
We close this section by observing that linear possibility regression gives triangular fuzzy numbers for Y, the
dependent variable. It is like doing interval estimation, or getting a regression band. Readers who are seriously
interested in this topic should refer to Terano, et al. (see references).
Linear Possibility Regression Model 399


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Section II: Fuzzy Control

This section discusses the fuzzy logic controller (FLC), its application and design. Fuzzy control is used in a
variety of machines and processes today, with widespread application especially in Japan. A few of the
applications in use today are in the list in Table 16.10, adapted from Yan, et al.
Table 16.10 Applications of Fuzzy Logic Controllers (FLCs) and Functions Performed
ApplicationFLC function(s)
Video camcorderDetermine best focusing and lighting when there is movement in the picture
Washing machineAdjust washing cycle by judging the dirt, size of the load, and type of fabric
TelevisionAdjust brightness, color, and contrast of picture to please viewers
Motor controlImprove the accuracy and range of motion control under unexpected conditions
Subway trainIncrease the stable drive and enhance the stop accuracy by evaluating the passenger traffic
conditions. Provide a smooth start and smooth stop.
Vacuum cleanerAdjust the vacuum cleaner motor power by judging the amount of dust and dirt and the floor
characteristics
Hot water heaterAdjust the heating element power according to the temperature and the quantity of water
being used
Helicopter controlDetermine the best operation actions by judging human instructions and the flying
conditions including wind speed and direction
Designing a Fuzzy Logic Controller

A fuzzy logic controller diagram was shown in Chapter 3. Let us redraw it now and discuss a design example.
Refer to Figure 16.2. For the purpose of discussion, let us assume that this FLC controls a hot water heater.
The hot water heater has a knob, HeatKnob(0−10) on it to control the heating element power, the higher the
value, the hotter it gets, with a value of 0 indicating the heating element is turned off. There are two sensors in
the hot water heater, one to tell you the temperature of the water (TempSense), which varies from 0 to 125° C,
and the other to tell you the level of the water in the tank (LevelSense), which varies from 0 = empty to 10 =
full. Assume that there is an automatic flow control that determines how much cold water (at temperature 10°
C) flows into the tank from the main water supply; whenever the level of the water gets below 40, the flow
control turns on, and turns off when the level of the water gets above 95.
Figure 16.2 Fuzzy control of a water heater.
The design objective can be stated as:
Keep the water temperature as close to 80° C as possible, in spite of changes in the water flowing out of
the tank, and cold water flowing into the tank.
Section II: Fuzzy Control 400

Step One: Defining Inputs and Outputs for the FLC
The range of values that inputs and outputs may take is called the universe of discourse. We need to define the
universe of discourse for all of the inputs and outputs of the FLC, which are all crisp values. Table 16.11
shows the ranges:
Table 16.11 Universe of Discourse for Inputs and Outputs for FLC
NameInput/OutputMinimum valueMaximum value
LevelSenseI010
HeatKnobO010
TempSenseI0125
Step Two: Fuzzify the Inputs
The inputs to the FLC are the LevelSense and the TempSense. We can use triangular membership functions to
fuzzify the inputs, just as we did in Chapter 3, when we constructed the fuzzifier program. There are some
general guidelines you can keep in mind when you determine the range of the fuzzy variables as related to the
crisp inputs (adapted from Yan, et al.):
1. Symmetrically distribute the fuzzified values across the universe of discourse.

2. Use an odd number of fuzzy sets for each variable so that some set is assured to be in the middle.
The use of 5 to 7 sets is fairly typical.
3. Overlap adjacent sets (by 15% to 25% typically) .
Both the input variables LevelSense and TempSense are restricted to positive values. We use the following
fuzzy sets to describe them:
XSmall, Small, Medium, Large, XLarge
In Table 16.12 and Figure 16.3, we show the assignment of ranges and triangular fuzzy membership functions
for LevelSense. Similarly, we assign ranges and triangular fuzzy membership functions for TempSense in
Table 16.13 and Figure 16.4. The optimization of these assignments is often done through trial and error for
achieving optimum performance of the FLC.
Table 16.12 Fuzzy Variable Ranges for LevelSense

Crisp Input RangeFuzzy Variable
0–2XSmall
1.5–4Small
3–7Medium
6–8.5Large
7.5–10XLarge
Figure 16.3 Fuzzy membership functions for LevelSense.
Table 16.13 Fuzzy Variable Ranges for TempSense
Step One: Defining Inputs and Outputs for the FLC 401

0–20XSmall
10–35Small
30–75Medium
60–95Large
85–125XLarge
Figure 16.4 Fuzzy membership functions for TempSense.
Step One: Defining Inputs and Outputs for the FLC 402

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Step Three: Set Up Fuzzy Membership Functions for the Output(s)
In our example, we have just one output, which is the HeatKnob. We need to assign fuzzy memberships to
this variable just as we did for the inputs. This is shown in Table 16.14 and Figure 16.5. We use different
variable names to make the example clearer later on.
Table 16.14 Fuzzy Variable Ranges for HeatKnob

0–2VeryLittle
1.5–4ALittle
3–7AGoodAmount
6–8.5ALot
7.5–10AWholeLot
Figure 16.5 Fuzzy membership functions for the output HeatKnob.
Step Four: Create a Fuzzy Rule Base
Now that you have the inputs and the output defined in terms of fuzzy variables, you need only specify what
actions to take under what conditions; that is, you need to construct a set of rules that describe the operation of
the FLC. These rules usually take the form of IF–THEN rules and can be obtained from a human expert
(heuristics), or can be supplied from a neural network that infers the rules from behavior of the system. We
mentioned this idea in Chapter 3.
Let us construct a rule base for our design example. For the two inputs, we define the matrix shown in Table
16.15. Our heuristic guidelines in determining this matrix are the following statements and their converses:
1. When the temperature is low, the HeatKnob should be set higher than when the temperature is
high.
2. When the volume of water is low, the HeatKnob does not need to be as high as when the volume
of water is high.
NOTE: In FLCs, we do not need to specify all the boxes in the matrix. That is perfectly
fine. No entry signifies that no action is taken, for example, in the column for
SenseTemp=XL, no action is required since the temperature is already at or above the target
temperature.
Step Three: Set Up Fuzzy Membership Functions for the Output(s) 403
Table 16.15 Fuzzy Rule Base for the Example Design, Output HeatKnob
SenseTemp−> Sense LevelXSSMLXL
\/
XSAGoodAmountALittleVeryLittle
SALotAGoodAmountVeryLittleVeryLittle
MAWholeLotALotAGoodAmountVeryLittle
LAWholeLotALotALotALittle
XLAWholeLotALotALotAGoodAmount
Let us examine a couple of typical entries in the table: For SenseLevel = Medium (M) and SenseTemp =
XSmall (XS), the output is HeatKnob = AWholeLot. Now for the same temperature, as the water level rises,
the setting on HeatKnob also should rise to compensate for the added volume of water. You can see that for
SenseLevel = Large(L), and SenseTemp = XSmall(XS), the output is HeatKnob = AWholeLot. You can
verify that the rest of the table is created by similar reasoning.
Creating IF–THEN Rules

We can now translate the table entries into IF − THEN rules. We take these directly from Table 16.15:
1. IF SenseTemp IS XSmall AND SenseLevel IS XSmall THEN SET HeatKnob TO

AGoodAmount
2. IF SenseTemp IS XSmall AND SenseLevel IS Small THEN SET HeatKnob TO ALot
3. IF SenseTemp IS XSmall AND SenseLevel IS Medium THEN SET HeatKnob TO AWholeLot
4. IF SenseTemp IS XSmall AND SenseLevel IS Large THEN SET HeatKnob TO AWholeLot
5. IF SenseTemp IS XSmall AND SenseLevel IS XLarge THEN SET HeatKnob TO AWholeLot
6. IF SenseTemp IS Small AND SenseLevel IS XSmall THEN SET HeatKnob TO ALittle
7. IF SenseTemp IS Small AND SenseLevel IS Small THEN SET HeatKnob TO AGoodAmount
8. IF SenseTemp IS Small AND SenseLevel IS Medium THEN SET HeatKnob TO ALot
9. IF SenseTemp IS Small AND SenseLevel IS Large THEN SET HeatKnob TO ALot
10. IF SenseTemp IS Small AND SenseLevel IS XLarge THEN SET HeatKnob TO ALot
11. IF SenseTemp IS Medium AND SenseLevel IS XSmall THEN SET HeatKnob TO VeryLittle
12. IF SenseTemp IS Medium AND SenseLevel IS Small THEN SET HeatKnob TO VeryLittle
13. IF SenseTemp IS Medium AND SenseLevel IS Medium THEN SET HeatKnob TO
AGoodAmount
14. IF SenseTemp IS Medium AND SenseLevel IS Large THEN SET HeatKnob TO ALot
15. IF SenseTemp IS Medium AND SenseLevel IS XLarge THEN SET HeatKnob TO ALot
16. IF SenseTemp IS Large AND SenseLevel IS Small THEN SET HeatKnob TO VeryLittle
17. IF SenseTemp IS Large AND SenseLevel IS Medium THEN SET HeatKnob TO VeryLittle
18. IF SenseTemp IS Large AND SenseLevel IS Large THEN SET HeatKnob TO ALittle
19. IF SenseTemp IS Large AND SenseLevel IS XLarge THEN SET HeatKnob TO
AGoodAmount
Remember that the output and inputs to the fuzzy rule base are fuzzy variables. For any given crisp input
value, there may be fuzzy membership in several fuzzy input variables (determined by the fuzzification step).
And each of these fuzzy input variable activations will cause different fuzzy output cells to fire, or be
activated. This brings us to the final step, defuzzification of the output into a crisp value.
Step Three: Set Up Fuzzy Membership Functions for the Output(s) 404
Step Five: Defuzzify the Outputs
In order to control the HeatKnob, we need to obtain a crisp dial setting. So far, we have several of the
IF–THEN rules of the fuzzy rule base firing at once, because the inputs have been fuzzified. How do we
arrive at a single crisp output number ? There are actually several different strategies for this; we will consider
two of the most common, the center of area (COA) or centroid method, and the fuzzy Or method. The easiest
way to understand the process is with an example.
Step Five: Defuzzify the Outputs 405


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Assume that at a particular point in time, LevelSense = 7.0 and TempSense = 65. These are the crisp inputs
directly from the sensors. With fuzzification (refer to Chapter 3 for a review), assume that you get the
following fuzzy memberships:
crisp input — LevelSense = 7.0

fuzzy outputs with membership values −
Medium: 0.4
Large: 0.6
all others : 0.0
crisp input — TempSense=65
fuzzy outputs with membership values −
Medium: 0.75
Large: 0.25
all others: 0.0
This results in four rules firing:
1. TempSense = Medium (0.75) AND LevelSense = Medium (0.4)

2. TempSense = Large (0.25) AND LevelSense = Medium (0.4)
3. TempSense = Medium (0.75) AND LevelSense = Large (0.6)
4. TempSense = Large (0.25) AND LevelSense = Large (0.6)
First you must determine, for each of the AND clauses in the IF–THEN rules, what the output should be. This
is done by the conjunction or minimum operator. So for each of these rules you have the following firing
strengths:
1. (0.75) ¥ (0.4) = 0.4

2. (0.25) ¥ (0.4) = 0.25
3. (0.75) ¥ (0.6) = 0.6
4. (0.25) ¥ (0.6) = 0.25
By using the fuzzy rule base and the strengths assigned previously, we find the rules recommend the
following output values (with strengths) for HeatKnob:
1. AGoodAmount (0.4)
2. VeryLittle (0.25)
3. ALot (0.6)

4. ALittle (0.25)
Now we must combine the recommendations to arrive at a single crisp value. First, we will use the fuzzy Or
method of defuzzification. Here we use a disjunction or maximum operator to combine the values. We obtain
the following:
(0.4) ¦ (0.25) ¦ (0.6) ¦ (0.25) = 0.6
The crisp output value for HeatKnob would then be this membership value multiplied by the range of the
output variable, or (0.6) (10−0) = 6.0.
Another way of combining the outputs is with the centroid method. With the centroid method, there are two
variations, the overlap composition method and the additive composition method. To review, we have the
following output values and strengths.
1. AGoodAmount (0.4)
2. VeryLittle (0.25)
3. ALot (0.6)
4. ALittle (0.25)
We use the strength value and fill in the particular triangular membership function to that strength level. For
example, for the first rule, we fill the triangular membership function, AGoodAmount to the 0.4 level. We
then cut off the top of the triangle (above 0.4). Next we do the same for the other rules. Finally we align the
truncated figures as shown in Figure 16.6 and combine them according to the overlap composition method or
the additive composition method (Kosko). You can see the difference in these two composition methods in
Figure 16.6. The overlap method simply superimposes all of the truncated triangles onto the same area. You
can lose information with this method. The additive method adds the geometrical figures on top of each other.
Figure 16.6 Defuzzification with the Centroid approach: Overlap and Additive composition.
In order to get a crisp value, you take the centroid or center of gravity, of the resulting geometrical figure. Let
us do this for the overlap method figure. The centroid is a straight edge that can be placed through the figure
to have it perfectly balanced; there is equal area of the figure on either side of the straight edge, as shown in
Figure 16.7. Splitting up the geometry into pieces and summing all area contributions on either side of the
centroid, we get a value of 5.2 for this example. This is already in terms of the crisp output value range:
HeatKnob = 5.2
Figure 16.7 Finding the centroid.

This completes the design for the simple example we chose. We conclude with a list of the advantages and
disadvantages of FLCs.
Advantages and Disadvantages of Fuzzy Logic Controllers

The following list adapted from McNeill and Thro shows the advantages and disadvantages to FLCs for
control systems as compared to more traditional control systems.
Advantages:
• Relates input to output in linguistic terms, easily understood

• Allows for rapid prototyping because the system designer doesn’t need to know everything about
the system before starting
• Cheaper because they are easier to design
• Increased robustness
• Simplify knowledge acquisition and representation
• A few rules encompass great complexity
• Can achieve less overshoot and oscillation
• Can achieve steady state in a shorter time interval
Disadvantages:
• Hard to develop a model from a fuzzy system

• Require more fine tuning and simulation before operational
• Have a stigma associated with the word fuzzy (at least in the Western world); engineers and most
other people are used to crispness and shy away from fuzzy control and fuzzy decision making
Summary
Fuzzy logic applications are many and varied. You got an overview of the different applications areas that
exist for fuzzy logic, from the control of washing machines to fuzzy logic based cost benefit analysis. Further
you got details on two application domains: fuzzy databases and fuzzy control.
This chapter dealt with extending database models to accommodate fuzziness in data attribute values and in
queries as well. You saw fuzzy relations; in particular, similarity and resemblance relations, and similarity
classes were reviewed. You found how possibility distributions help define fuzzy databases. You also learned
what fuzzy events are and how to calculate fuzzy means, fuzzy variances, and fuzzy conditional expectations.
Concepts related to linear possibility regression model were presented.
The chapter presented the design of a simple fuzzy logic control (FLC) system to regulate the temperature of
water in a hot water heater. The components of the FLC were discussed along with design procedures. The
advantages of FLC design include rapid prototyping ability and the capability to solve very nonlinear control
problems without knowing details of the nonlinearities.
Advantages and Disadvantages of Fuzzy Logic Controllers 408


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Chapter 17
Further Applications
Introduction
In this chapter, we present the outlines of some applications of neural networks and fuzzy logic. Most of the
applications fall into a few main categories according to the paradigms they are based on. We offer a sampling
of topics of research as found in the current literature, but there are literally thousands of applications of
neural networks and fuzzy logic in science, technology and business with more and more applications added
as time goes on.
Some applications of neural networks are for adaptive control. Many such applications benefit from adding
fuzziness also. Steering a car or backing up a truck with a fuzzy controller is an example. A large number of
applications are based on the backpropagation training model. Another category of applications deals with
classification. Some applications based on expert systems are augmented with a neural network approach.
Decision support systems are sometimes designed this way. Another category is made up of optimizers, whose
purpose is to find the maximum or the minimum of a function.
NOTE: You will find other neural network applications related to finance presented toward
the end of Chapter 14.
Computer Virus Detector

IBM Corporation has applied neural networks to the problem of detecting and correcting computer viruses.
IBM’s AntiVirus program detects and eradicates new viruses automatically. It works on boot−sector types of
viruses and keys off of the stereotypical behaviors that viruses usually exhibit. The feedforward
backpropagation neural network was used in this application. New viruses discovered are used in the training
set for later versions of the program to make them “smarter.” The system was modeled after knowledge about
the human immune system: IBM uses a decoy program to “attract” a potential virus, rather than have the virus
attack the user’s files. These decoy programs are then immediately tested for infection. If the behavior of the
decoy program seems like the program was infected, then the virus is detected on that program and removed
wherever it’s found.
Mobile Robot Navigation

C. Lin and C. Lee apply a multivalued Boltzmann machine, modeled by them, using an artificial magnetic
field approach. They define attractive and repulsive magnetic fields, corresponding to goal position and
obstacle, respectively. The weights on the connections in the Boltzmann machine are none other than the
magnetic fields.
Chapter 17 Further Applications 409

They divide a two−dimensional traverse map into small grid cells. Given the goal cell and obstacle cells, the
problem is to navigate the two−dimensional mobile robot from an unobstructed cell to the goal quickly,
without colliding with any obstacle. An attracting artificial magnetic field is built for the goal location. They
also build a repulsive artificial magnetic field around the boundary of each obstacle. Each neuron, a grid cell,
will point to one of its eight neighbors, showing the direction for the movement of the robot. In other words,
the Boltzmann machine is adapted to become a compass for the mobile robot.
A Classifier
James Ressler and Marijke Augusteijn study the use of neural networks to the problem of weapon to target
assignment. The neural network is used as a filter to remove unfeasible assignments, where feasibility is
determined in terms of the weapon’s ability to hit a given target if fired at a specific instant. The large number
of weapons and threats along with the limitation on the amount of time lend significance to the need for
reducing the number of assignments to consider.
The network’s role here is classifier, as it needs to separate the infeasible assignments from the feasible ones.
Learning has to be quick, and so Ressler and Augusteijn prefer to use an architecture called the
cascade−correlation learning architecture, over backpropagation learning. Their network is dynamic in that
the number of hidden layer neurons is determined during the training phase. This is part of a class of
algorithms that change the architecture of the network during training.
A Two−Stage Network for Radar Pattern Classification

Mohammad Ahmadian and Russell Pimmel find it convenient to use a multistage neural network
configuration, a two−stage network in particular, for classifying patterns. The patterns they study are
geometrical features of simulated radar targets.
Feature extraction is done in the first stage, while classification is done in the second. Moreover, the first stage
is made up of several networks, each for extracting a different estimable feature. Backpropagation is used for
learning in the first stage. They use a single network in the second stage. The effect of noise is also studied.
A Classifier 410

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Crisp and Fuzzy Neural Networks for Handwritten Character Recognition

Paul Gader, Magdi Mohamed, and Jung−Hsien Chiang combine a fuzzy neural network and a crisp neural
network for the recognition of handwritten alphabetic characters. They use backpropagation for the crisp
neural network and a clustering algorithm called K−nearest neighbor for the fuzzy network. Their
consideration of a fuzzy network in this study is prompted by their belief that if some ambiguity is possible in
deciphering a character, such ambiguity should be accurately represented in the output. For example, a
handwritten “u” could look like a “v” or “u.” If present, the authors feel that this ambiguity should be
translated to the classifier output.
Feature extraction was accomplished as follows: character images of size 24x16 pixels were used. The first
stage of processing extracted eight feature images from the input image, two for each direction (north,
northeast, northwest, and east). Each feature image uses an integer at each location that represents the length
of the longest bar that fits at that point in that direction. These are referred to as bar features. Next 8x8
overlapping zones are used on the feature images to derive feature vectors. These are made by taking the
summed values of the values in a zone and dividing this by the maximum possible value in the zone. Each
feature image results in a 15,120 element feature vectors.
Data was obtained from the U.S. Postal Office, consisting of 250 characters. Results showed 97.5% and
95.6% classification rates on training and test sets, respectively, for the neural network. The fuzzy network
resulted in 94.7% and 93.8% classification rates, where the desired output for many characters was set to
ambiguous.
Noise Removal with a Discrete Hopfield Network

Arun Jagota applies what is called a HcN, a special case of a discrete Hopfield network, to the problem of
recognizing a degraded printed word. HcN is used to process the output of an Optical Character Recognizer,
by attempting to remove noise. A dictionary of words is stored in the HcN and searched.
Object Identification by Shape

C. Ganesh, D. Morse, E. Wetherell, and J. Steele used a neural network approach to an object identification
system, based on the shape of an object and independent of its size. A two−dimensional grid of ultrasonic data
represents the height profile of an object. The data grid is compressed into a smaller set that retains the
essential features. Backpropagation is used. Recognition on the order of approximately 70% is achieved.
Detecting Skin Cancer

F. Ercal, A. Chawla, W. Stoecker, and R. Moss study a neural network approach to the diagnosis of malignant
melanoma. They strive to discriminate tumor images as malignant or benign. There are as many as three
categories of benign tumors to be distinguished from malignant melanoma. Color images of skin tumors are
Crisp and Fuzzy Neural Networks for Handwritten Character Recognition 411
used in the study. Digital images of tumors are classified. Backpropagation is used. Two approaches are taken
to reduce training time. The first approach involves using fewer hidden layers, and the second involves
randomization of the order of presentation of the training set.
EEG Diagnosis
Fred Wu, Jeremy Slater, R. Eugene Ramsay, and Lawrence Honig use a feedforward backpropagation neural
network as a classifier in EEG diagnosis. They compare the performance of the neural network classifier to
that of a nearest neighbor classifier. The neural network classifier shows a classifier accuracy of 75% for
Multiple Sclerosis patients versus 65% for the nearest neighbor algorithm.
Time Series Prediction with Recurrent and Nonrecurrent Networks

Sathyanarayan Rao and Sriram Sethuraman take a recurrent neural network and a feedforward network and
train then in parallel. A recurrent neural network has feedback connections from the output neurons back to
input neurons to model the storage of temporal information. A modified backpropagation algorithm is used
for training the recurrent network, called the real−time recurrent learning algorithm. They have the recurrent
neural network store past information, and the feedforward network do the learning of nonlinear dependencies
on the current samples. They use this scheme because the recurrent network takes more than one time period
to evaluate its output, whereas the feedforward network does not. This hybrid scheme overcomes the latency
problem for the recurrent network, providing immediate nonlinear evaluation from input to output.
Security Alarms
Deborah Frank and J. Bryan Pletta study the application of neural networks for alarm classification based on
their operation under varying weather conditions. Performance degradation of a security system when the
environment changes is a cause for losing confidence in the system itself. This problem is more acute with
portable security systems.
They investigated the problem using several networks, ranging from backpropagation to learning vector
quantization. Data was collected using many scenarios, with and without the coming of an intruder, which can
be a vehicle or a human.
They found a 98% probability of detection and 9% nuisance alarm rate over all weather conditions.
EEG Diagnosis 412


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Circuit Board Faults

Anthony Mason, Nicholas Sweeney, and Ronald Baer studied the neural network approach in two laboratory
experiments and one field experiment, in diagnosing faults in circuit boards.
Test point readings were expressed as one vector. A fault vector was also defined with elements representing
possible faults. The two vectors became a training pair. Backpropagation was used.
Warranty Claims
Gary Wasserman and Agus Sudjianto model the prediction of warranty claims with neural networks. The
nonlinearity in the data prompted this approach.
The motivation for the study comes from the need to assess warranty costs for a company that offers extended
warranties for its products. This is another application that uses backpropagation. The architecture used was
2−10−1.
Writing Style Recognition

J. Nellis and T. Stonham developed a neural network character recognition system that adapts dynamically to
a writing style.
They use a hybrid neural network for hand−printed character recognition, that integrates image processing and
neural network architectures. The neural network uses random access memory (RAM) to model the
functionality of an individual neuron. The authors use a transform called the five−way image processing
transform on the input image, which is of size 32x32 pixels. The transform converts the high spatial frequency
data in a character into four low frequency representations. What they achieve by this are position invariance,
and a ratio of black to white pixels approaching 1, rotation invariance, and capability to detect and correct
breaks within characters. The transformed data are input to the neural network that is used as a classifier and
is called a discriminator.
A particular writing style that has less variability and therefore fewer subclasses is needed to classify the style.
Network size will also reduce confusion, and conflicts lessen.
Commercial Optical Character Recognition

Optical character recognition (OCR) is one of the most successful commercial applications of neural
networks. Caere Corporation brought out its neural network product in 1992, after studying more than
100,000 examples of fax documents. Caere’s AnyFax technology combines neural networks with expert
systems to extract character information from Fax or scanned images. Calera, another OCR vendor, started
using neural networks in 1984 and also benefited from using a very large (more than a million variations of
Circuit Board Faults 413

alphanumeric characters) training set.
ART−EMAP and Object Recognition

A neural network architecture called ART−EMAP (Gail Carpenter and William Ross) integrates Adaptive
Resonance Theory (ART) with spatial and temporal evidence integration for predictive mapping (EMAP).
The result is a system capable of complex 3−D object recognition. A vision system that samples
two−dimensional perspectives of a three−dimensional object is created that results in 98% correct recognition
with an average of 9.2 views presented on noiseless test data, and 92% recognition with an average of 11.2
views presented on noisy test data. The ART−EMAP system is an extension of ARTMAP, which is a neural
network architecture that performs supervised learning of recognition categories and multidimensional maps
in response to input vectors. A fuzzy logic extension of ARTMAP is called Fuzzy ARTMAP, which
incorporates two fuzzy modules in the ART system.
Summary
A sampling of current research and commercial applications with neural networks and fuzzy logic technology
is presented. Neural networks are applied toward a wide variety of problems, from aiding medical diagnosis to
detecting circuit faults in printed circuit board manufacturing. Some of the problem areas where neural
networks and fuzzy logic have been successfully applied are:
• Filtering
• Image processing
• Intelligent control
• Machine vision
• Motion analysis
• Optimization
• Pattern recognition
• Prediction
• Time series analysis
• Speech synthesis
• Machine learning and robotics
• Decision support systems
• Classification
• Data compression
• Functional approximation
The use of fuzzy logic and neural networks in software and hardware systems can only increase!
ART−EMAP and Object Recognition 414


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Table of Contents
Appendix A
Compiling Your Programs
All of the programs included in the book have been compiled and tested on Turbo C++, Borland C++, and
Microsoft C++/Visual C++ with either the small or medium memory model. You should not have any
problems in using other compilers, since standard I/O routines are used. Your target should be a DOS
executable. With the backpropagation simulator of Chapters 7, 13, and 14 you may run into a memory
shortage situation. You should unload any TSR (Terminate and Stay Resident) programs and/or choose
smaller architectures for your networks. By going to more hidden layers with fewer neurons per layer, you
may be able to reduce the overall memory requirements.
The programs in this book make heavy use of floating point calculations, and you should compile your
programs to take advantage of a math coprocessor, if you have one installed in your computer.
The organization of files on the accompanying diskette are according to chapter number. You will find
relevant versions of files in the corresponding chapter directory.
Most of the files are self−contained, or include other needed files in them. You will not require a makefile to
build the programs. Load the main file for example for backpropagation, the backprop.cpp file, into the
development environment editor for your compiler and build a .exe file. That’s it!
Table of Contents
Appendix A Compiling Your Programs 415


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Table of Contents
Appendix B
Mathematical Background
Dot Product or Scalar Product of Two Vectors
Given vectors U and V, where U = (u1, …, un) and V = (v1, …, vn), their dot product or scalar product is U " V
= u1v1 +… + unvn = £ ui vi.
Matrices and Some Arithmetic Operations on Matrices

A real matrix is a rectangular array of real numbers. A matrix with m rows and n columns is referred to as an
mxn matrix. The element in the ith row and jth column of the matrix is referred to as the ij element of the
matrix and is denoted by aij.
The transpose of a matrix M is denoted by MT. The element in the ith row and jth column of MT is the same as
the element of M in its jth row and ith column. MT is obtained from M by interchanging the rows and columns
of M. For example, if
2 7 −3 2 4
M = , then MT = 7 0
4 0 9 −3 9
If X is a vector with m components, x1, …, xm, then it can be written as a column vector with components
listed one below another. It can be written as a row vector, X = (x1, …, xm). The transpose of a row vector is
the column vector with the same components, and the transpose of a column vector is the corresponding row
vector.
The addition of matrices is possible if they have the same size, that is, the same number of rows and same
number of columns. Then you just add the ij elements of the two matrices to get the ij elements of the sum
matrix. For example,
3 −4 5 5 2 −3 8 −2 2
+ =
2 3 7 6 0 4 8 3 11
Multiplication is defined for a given pair of matrices, only if a condition on their respective sizes is satisfied.
Then too, it is not a commutative operation. This means that if you exchange the matrix on the left with the
matrix on the right, the multiplication operation may not be defined, and even if it is, the result may not be the
same as before such an exchange.
The condition to be satisfied for multiplying the matrices A, B as AB is, that the number of columns in A is
equal to the number of rows in B. Then to get the ij element of the product matrix AB, you take the ith row of
Appendix B Mathematical Background 416

A as one vector and the jth column of B as a second vector and do a dot product of the two. For example, the
two matrices given previously to illustrate the addition of two matrices are not compatible for multiplication
in whichever order you take them. It is because there are three columns in each, which is different from the
number of rows, which is 2 in each. Another example is given as follows.
3 −4 5 5 6
Let A = and B = 2 0
2 3 7 −3 4
Then AB and BA are both defined, AB is a 2x2 matrix, whereas BA is 3x3.
−8 38 27 −2 67
Also AB = and BA = 6 −8 10
−5 40 −1 24 13
Lyapunov Function
A Lyapunov function is a function that decreases with time, taking on non−negative values. It is used to
correspond between the state variables of a system and real numbers. The state of the system changes as time
changes, and the function decreases. Thus, the Lyapunov function decreases with each change of state of the
system.
We can construct a simple example of a function with the property of decreasing with each change of state as
follows. Suppose a real number, x, represents the state of a dynamic system at time t. Also suppose that x is
bounded for any t by a positive real number M. That means x is less than M for every value of t.
Then the function,
f(x,t) = exp(−|x|/(M+|x|+t))
is non−negative and decreases with increasing t.
Local Minimum
A function f(x) is defined to have a local minimum at y, with a value z, if
f(y) = z, and f(x) e z, for each x, such that there exists

a positive real number h such that y – h d x d y + h.
In other words, there is no other value of x in a neighborhood of y, where the value of the function is smaller
than z.
There can be more than one local minimum for a function in its domain. A Step function (with a graph
resembling a staircase) is a simple example of a function with an infinite number of points in its domain with
local minima.
Global Minimum
A function f(x) is defined to have a global minimum at y, with a value z, if
f(y) = z, and f(x) e z, for each x in the domain of

the function f.
Global Minimum 417

In other words, there is no other value of x in the domain of the function f, where the value of the function is
smaller than z. Clearly, a global minimum is also a local minimum, but a local minimum may not be a global
minimum.
There can be more than one global minimum for a function in its domain. The trigonometric function f(x) =
sinx is a simple example of a function with an infinite number of points with global minima. You may recall
that sin(3À/ 2), sin (7À/ 2), and so on are all –1, the smallest value for the sine function.
Kronecker Delta Function

The Kronecker delta function is a function of two variables. It has a value of 1 if the two arguments are equal,
and 0 if they are not. Formally,
1 if x = y
´(x,y)=
0 if x `y
Gaussian Density Distribution

The Gaussian Density distribution, also called the Normal distribution, has a density function of the following
form. There is a constant parameter c, which can have any positive value.
Table of Contents
Global Minimum 418


by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Table of Contents
References
Ahmadian, Mohamad, and Pimmel, Russell, “Recognizing Geometrical Features of Simulated Targets
with Neural Networks,” Conference Proceedings of the 1992 Artificial Neural Networks in
Engineering Conference, V.3, pp. 409–411.
Aleksander, Igor, and Morton, Helen, An Introduction to Neural Computing, Chapman and Hall,
London, 1990.
Aiken, Milam, “Forecasting T−Bill Rates with a Neural Net,” Technical Analysis of Stocks and
Commodities, May 1995, Technical Analysis Inc., Seattle.
Anderson, James, and Rosenfeld, Edward, eds., Neurocomputing: Foundations of Research, MIT
Press, Cambridge, MA, 1988.
Anzai, Yuichiro, Pattern Recognition and Machine Learning, Academic Press, Englewood Cliffs, NJ,
1992.
Azoff, E. Michael, Neural Network Time Series Forecasting of Financial Markets, John Wiley &
Sons, New York, 1994.
Bauer, Richard J., Genetic Algorithms and Investment Strategies, John Wiley & Sons, New York,
1994.
Booch, Grady, Object Oriented Design with Applications, Benjamin−Cummings, Redwood City, CA,
1991.
Carpenter, Gail A., and Ross, William D., “ART−EMAP: A Neural Network Architecture for Object
Recognition by Evidence Accumulation,” IEEE Transactions on Neural Networks, Vol. 6., No. 4,
July 1995, pp. 805–818.
Colby, Robert W., and Meyers, Thomas A., The Encyclopedia of Technical Market Indicators,
Business One Irwin, Homewood, IL, 1988.
Collard, J. E., “Commodity Trading with a Three Year Old,” Neural Networks in Finance and
Investing, pp. 411–420, Probus Publishing, Chicago, 1993.
Cox, Earl, The Fuzzy Systems Handbook, Academic Press, Boston, 1994.
Dagli, Cihan, et al., eds., Intelligent Engineering Systems through Artificial Neural Networks, Volume
2, ASME Press, New York, 1992.
Davalo, Eric, and Naim, Patrick, Neural Networks, MacMillan, New York, 1991.
Ercal, F., et al., “Diagnosing Malignant Melanoma Using a Neural Network,” Conference
Proceedings of the 1992 Artificial Neural Networks in Engineering Conference, V.3, pp. 553–555.
Frank, Deborah, and Pletta, J. Bryan, “Neural Network Sensor Fusion for Security Applications,”
Conference Proceedings of the 1992 Artificial Neural Networks in Engineering Conference, V.3, pp.
745–748.
Freeman, James A., and Skapura, David M., Neural Networks Algorithms, Applications, and
Programming Techniques, Addison−Wesley, Reading, MA, 1991.
Gader, Paul, et al., “Fuzzy and Crisp Handwritten Character Recognition Using Neural Networks,”
Conference Proceedings of the 1992 Artificial Neural Networks in Engineering Conference, V.3, pp.
421–424.
Ganesh, C., et al., “A Neural Network−Based Object Identification System,” Conference Proceedings
of the 1992 Artificial Neural Networks in Engineering Conference, V.3, pp. 471–474.
References 419
Glover, Fred, ORSA CSTS Newsletter Vol 15, No 2, Fall 1994.

Goldberg, David E., Genetic Algorithms in Search, Optimization and Machine Learning,
Addison−Wesley, Reading, MA, 1989.
Grossberg, Stephen, et al., Introduction and Foundations, Lecture Notes, Neural Network Courses and
Conference, Boston University, May 1992.
Hammerstrom, Dan, “Neural Networks at Work,” IEEE Spectrum, New York, June 1993.
Hertz, John, Krogh, Anders, and Palmer, Richard, Introduction to the Theory of Neural Computation,
Addison−Wesley, Reading, MA, 1991.
Jagota, Arun, “Word Recognition with a Hopfield−Style Net,” Conference Proceedings of the 1992
Artificial Neural Networks in Engineering Conference, V.3, pp. 445–448.
Johnson, R. Colin, “Accuracy Moves OCR into the Mainstream,” Electronic Engineering Times,
CMP Publications, Manhasset, NY, January 16, 1995.
Johnson, R. Colin, “Making the Neural−Fuzzy Connection,” Electronic Engineering Times, CMP
Publications, Manhasset, NY, September 27, 1993.
Johnson, R. Colin, “Neural Immune System Nabs Viruses,” Electronic Engineering Times, CMP
Publications, Manhasset, NY, May 8, 1995.
Jurik, Mark, “The Care and Feeding of a Neural Network,” Futures Magazine, Oster
Communications, Cedar Falls, IA, October 1992.
Kimoto, Takashi, et al., “Stock Market Prediction System with Modular Neural Networks,” Neural
Networks in Finance and Investing, pp. 343–356, Probus Publishing, Chicago, 1993.
Kline, J., and Folger, T.A., Fuzzy Sets, Uncertainty and Information, Prentice Hall, New York, 1988.
Konstenius, Jeremy G., “Trading the S&P with a Neural Network,” Technical Analysis of Stocks and
Commodities, October 1994, Technical Analysis Inc., Seattle.
Kosaka, M., et al., “Applications of Fuzzy Logic/Neural Network to Securities Trading Decision
Support System,” Conference Proceedings of the 1991 IEEE International Conference on Systems,
Man and Cybernetics, V.3, pp. 1913–1918.
Kosko, Bart, and Isaka, Satoru, Fuzzy Logic, Scientific American, New York, July 1993.
Kosko, Bart, Neural Networks and Fuzzy Systems: A Dynamical Systems Approach to Machine
Intelligence, Prentice−Hall, New York, 1992.
Laing, Jonathan, “New Brains: How Smart Computers are Beating the Stock Market,” Barron’s,
February 27, 1995.
Lederman, Jess, and Klein, Robert, eds., Virtual Trading, Probus Publishing, Chicago, 1995.
Lin, C.T. and Lee, C.S.G, “A Multi−Valued Boltzman Machine”, IEEE Transactions on Systems,
Man, and Cybernetics, Vol. 25, No. 4, April 1995 pp. 660−668.
MacGregor, Ronald J., Neural and Brain Modeling, Academic Press, Englewood Cliffs, NJ, 1987.
Mandelman, Avner, “The Computer’s Bullish! A Money Manager’s Love Affair with Neural
Network Programs,” Barron’s, December 14, 1992.
Maren, Alianna, Harston, Craig, and Pap, Robert, Handbook of Neural Computing Applications,
Academic Press, Englewood Cliffs, NJ, 1990.
Marquez, Leorey, et al., “Neural Network Models as an Alternative to Regression,” Neural Networks
in Finance and Investing, pp. 435–449, Probus Publishing, Chicago, 1993.
Mason, Anthony et al., “Diagnosing Faults in Circuit Boards—A Neural Net Approach,” Conference
Proceedings of the 1992 Artificial Neural Networks in Engineering Conference, V.3, pp. 839–843.
McNeill, Daniel, and Freiberger, Paul, Fuzzy Logic, Simon & Schuster, New York, 1993.
McNeill, F. Martin, and Thro, Ellen, Fuzzy Logic: A Practical Approach, Academic Press, Boston,
1994.
Murphy, John J., Intermarket Technical Analysis, John Wiley & Sons, New York, 1991.
Murphy, John J., Technical Analysis of the Futures Markets, NYIF Corp., New York, 1986.
Nellis, J., and Stonham, T.J., “A Neural Network Character Recognition System that Dynamically
Adapts to an Author’s Writing Style,” Conference Proceedings of the 1992 Artificial Neural
Networks in Engineering Conference, V.3, pp. 975–979.
References 420
Peters, Edgar E., Fractal Market Analysis: Applying Chaos Theory to Investment and Economics,
John Wiley & Sons, New York, 1994.
Ressler, James L., and Augusteijn, Marijke F., “Weapon Target Assignment Accessibility Using
Neural Networks,” Conference Proceedings of the 1992 Artificial Neural Networks in Engineering
Conference, V.3, pp. 397–399.
Rao, Satyanaryana S,, and Sethuraman, Sriram, “A Modified Recurrent Learning Algorithm for
Nonlinear Time Series Prediction,” Conference Proceedings of the 1992 Artificial Neural Networks in
Engineering Conference, V.3, pp. 725–730.
Rao, Valluru, and Rao, Hayagriva, Power Programming Turbo C++, MIS:Press, New York, 1992.
Ruggiero, Murray, “How to Build an Artificial Trader,” Futures Magazine, Oster Communications,
Cedar Falls, IA, September 1994.
Ruggiero, Murray, “How to Build a System Framework,” Futures Magazine, Oster Communications,
Cedar Falls, IA, November 1994.
Ruggiero, Murray, “Nothing Like Net for Intermarket Analysis,” Futures Magazine, Oster
Communications, Cedar Falls, IA, May 1995.
Ruggiero, Murray, “Putting the Components before the System,” Futures Magazine, Oster
Communications, Cedar Falls, IA, October 1994.
Ruggiero, Murray, “Training Neural Networks for Intermarket Analysis,” Futures Magazine, Oster
Communications, Cedar Falls, IA, August 1994.
Rumbaugh, James, et al, Object−Oriented Modeling and Design, Prentice−Hall Inc, Englewood
Cliffs, New Jersey, 1991.
Sharda, Ramesh, and Patil, Rajendra, “A Connectionist Approach to Time Series Prediction: An
Empirical Test,” Neural Networks in Finance and Investing, pp. 451–463, Probus Publishing,
Chicago, 1993.
Sherry, Clifford, “Are Your Inputs Correlated?” Technical Analysis of Stocks and Commodities,
February 1995, Technical Analysis Inc., Seattle.
Simpson, Patrick K., Artificial Neural Systems: Foundations, Paradigms, Applications and
Implementations, Pergamon Press, London, 1990.
Soucek, Branko, and Soucek, Marina, Neural and Massively Parallel Computers: The Sixth
Generation, John Wiley & Sons, New York, 1988.
Stein, Jon, “The Trader’s Guide to Technical Indicators,” Futures Magazine, Oster Communications,
Cedar Falls, IA, August 1990.
Terano, Toshiro, et al., Fuzzy Systems Theory and Its Applications, Academic Press, Boston, 1993.
Trippi, Robert, and Turban, Efraim, eds., Neural Networks in Finance and Investing, Probus
Publishing, Chicago, 1993.
Wasserman, Gary S., and Sudjianto, Agus, Conference Proceedings of the 1992 Artificial Neural
Networks in Engineering Conference, V.3, pp. 901–904.
Wasserman, Philip D., Advanced Methods in Neural Computing, Van Nostrand Reinhold, New York,
1993.
Wasserman, Philip D., Neural Computing, Van Nostrand Reinhold, New York, 1989.
Wu, Fred, et al., “Neural Networks for EEG Diagnosis,” Conference Proceedings of the 1992
Artificial Neural Networks in Engineering Conference, V.3, pp. 559–565.
Yager, R., ed., Fuzzy Sets and Applications: Selected Papers by L.Z. Zadeh, Wiley−Interscience, New
York, 1987.
Yan, Jun, Ryan, Michael, and Power, James, Using Fuzzy Logic, Prentice−Hall, New York, 1994.
Yoon, Y., and Swales, G., “Predicting Stock Price Performance: A Neural Network Approach,”
Neural Networks in Finance and Investing, pp. 329–339, Probus Publishing, Chicago, 1993.
References 421
Table of Contents
References 422

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Table of Contents
Glossary
A
Activation
The weighted sum of the inputs to a neuron in a neural network.
Adaline
Adaptive linear element machine.
Theory developed by Grossberg and Carpenter for categorization of patterns, and to address the
stability–plasticity dilemma.
Algorithm
A step−by−step procedure to solve a problem.
Annealing
A process for preventing a network from being drawn into a local minimum.
ART
(Adaptive Resonance Theory) ART1 is the result of the initial development of this theory for binary
inputs. Further developments led to ART2 for analog inputs. ART3 is the latest.
Artificial neuron
The primary object in an artificial neural network to mimic the neuron activity of the brain. The
artificial neuron is a processing element of a neural network.
Associative memory
Activity of associating one pattern or object with itself or another.
Autoassociative
Making a correspondence of one pattern or object with itself.
B
Backpropagation
A neural network training algorithm for feedforward networks where the errors at the output layer are
propagated back to the layer before in learning. If the previous layer is not the input layer, then the
errors at this hidden layer are propagated back to the layer before.
BAM
Bidirectional Associative Memory network model.
Bias
A value added to the activation of a neuron.
Binary digit
A value of 0 or 1.
Bipolar value
A value of –1 or +1.
Boltzmann machine
Glossary 423
A neural network in which the outputs are determined with probability distributions. Trained and
operated using simulated annealing.
Anderson’s single−layer, laterally connected neural network model. It can work with inputs that have
noise in them or are incomplete.
C
Cauchy machine
Similar to the Boltzmann machine, except that a Cauchy distribution is used for probabilities.
Cognitron
The forerunner to the Neocognitron. A network developed to recognize characters.
Competition
A process in which a winner is selected from a layer of neurons by some criterion. Competition
suggests inhibition reflected in some connection weights being assigned a negative value.
Connection
A means of passing inputs from one neuron to another.
Connection weight
A numerical label associated with a connection and used in a weighted sum of inputs.
Constraint
A condition expressed as an equation or inequality, which has to be satisfied by the variables.
Convergence
Termination of a process with a final result.
Crisp
The opposite of fuzzy—usually a specific numerical quantity or value for an entity.
D
Delta rule
A rule for modification of connection weights, using both the output and the error obtained. It is also
called the LMS rule.
E
Energy function
A function of outputs and weights in a neural network to determine the state of the system, e.g.,
Lyapunov function.
Excitation
Providing positive weights on connections to enable outputs that cause a neuron to fire.
Exemplar
An example of a pattern or object used in training a neural network.
Expert system
A set of formalized rules that enable a system to perform like an expert.
F
FAM
Fuzzy Associative Memory network. Makes associations between fuzzy sets.
Feedback
The process of relaying information in the opposite direction to the original.
Glossary 424
Fit vector
A vector of values of degree of membership of elements of a fuzzy set.
Fully connected network
A neural network in which every neuron has connections to all other neurons.
Fuzzy
As related to a variable, the opposite of crisp. A fuzzy quantity represents a range of value as opposed
to a single numeric value, e.g., “hot” vs. 89.4°.
Fuzziness
Different concepts having an overlap to some extent. For example, descriptions of fair and cool
temperatures may have an overlap of a small interval of temperatures.
Fuzzy Associative Memory
A neural network model to make association between fuzzy sets.
Fuzzy equivalence relation
A fuzzy relation (relationship between fuzzy variables) that is reflexive, symmetric, and transitive.
Fuzzy partial order
A fuzzy relation (relationship between fuzzy variables) that is reflexive, antisymmetric, and transitive.
G
Gain
Sometimes a numerical factor to enhance the activation. Sometimes a connection for the same
purpose.
Generalized Delta rule
A rule used in training of networks such as backpropagation training where hidden layer weights are
modified with backpropagated error.
Global minimum
A point where the value of a function is no greater than the value at any other point in the domain of
the function.
H
Hamming distance
The number of places in which two binary vectors differ from each other.
Hebbian learning
A learning algorithm in which Hebb’s rule is used. The change in connection weight between two
neurons is taken as a constant times the product of their outputs.
Heteroassociative
Making an association between two distinct patterns or objects.
Hidden layer
An array of neurons positioned in between the input and output layers.
Hopfield network
A single layer, fully connected, autoassociative neural network.
I
Inhibition
The attempt by one neuron to diminish the chances of firing by another neuron.
Input layer
An array of neurons to which an external input or signal is presented.
Instar
Glossary 425
A neuron that has no connections going from it to other neurons.
L
Lateral connection
A connection between two neurons that belong to the same layer.
Layer
An array of neurons positioned similarly in a network for its operation.
Learning
The process of finding an appropriate set of connection weights to achieve the goal of the network
operation.
Linearly separable
Two subsets of a linear set having a linear barrier (hyperplane) between the two of them.
LMS rule
Least mean squared error rule, with the aim of minimizing the average of the squared error. Same as
the Delta rule.
Local minimum
A point where the value of the function is no greater than the value at any other point in its
neighborhood.
Long−term memory (LTM)
Encoded information that is retained for an extended period.
Lyapunov function
A function that is bounded below and represents the state of a system that decreases with every
change in the state of the system.
M
Madaline
A neural network in which the input layer has units that are Adalines. It is a multiple−Adaline.
Mapping
A correspondence between elements of two sets.
N
Neural network
A collection of processing elements arranged in layers, and a collection of connection edges between
pairs of neurons. Input is received at one layer, and output is produced at the same or at a different
layer.
Noise
Distortion of an input.
Nonlinear optimization
Finding the best solution for a problem that has a nonlinear function in its objective or in a constraint.
O
On center off surround
Assignment of excitatory weights to connections to nearby neurons and inhibitory weights to
connections to distant neurons.
Orthogonal vectors
Vectors whose dot product is 0.
Glossary 426
Outstar
A neuron that has no incoming connections.
P
Perceptron
A neural network for linear pattern matching.
Plasticity
Ability to be stimulated by new inputs and learn new mappings or modify existing ones.
R
Resonance
The responsiveness of two neurons in different layers in categorizing an input. An equilibrium in two
directions.
S
Saturation
A condition of limitation on the frequency with which a neuron can fire.
Self−organization
A process of partitioning the output layer neurons to correspond to individual patterns or categories,
also called unsupervised learning or clustering.
Short−term memory (STM)
The storage of information that does not endure long after removal of the corresponding input.
Simulated annealing
An algorithm by which changes are made to decrease energy or temperature or cost.
Stability
Convergence of a network operation to a steady−state solution.
Supervised learning
A learning process in which the exemplar set consists of pairs of inputs and desired outputs.
T
Threshold value
A value used to compare the activation of a neuron to determine if the neuron fires or not. Sometimes
a bias value is added to the activation of a neuron to allow the threshold value to be zero in
determining the neuron’s output.
Training
The process of helping in a neural network to learn either by providing input/output stimuli
(supervised training) or by providing input stimuli (unsupervised training), and allowing weight
change updates to occur.
U
Unsupervised learning
Learning in the absence of external information on outputs, also called self−organization or
clustering.
Glossary 427
V
Vigilance parameter
A parameter used in Adaptive Resonance Theory. It is used to selectively prevent the activation of a
subsystem in the network.
W
Weight
A number associated with a neuron or with a connection between two neurons, which is used in
aggregating the outputs to determine the activation of a neuron.
Table of Contents
Glossary 428

by Valluru B. Rao
ISBN: 1558515526 Pub Date: 06/01/95
Table of Contents
Index
A
ABAM see Adaptive Bi−directional Associative Memory
abstract class, 138
abstract data type, 22
accumulation−distribution, 402, 404,
activation, 3, 18, 89
zero, 182
Adaline, 81, 82, 102, 103, 112
adaptation, 77, 120
Adaptive Bi−directional Associative Memory, 181, 212
Competitive, 212
Differential
Competitive, 212
Hebbian, 212
adaptive filters, 120
adaptive linear element, 102
adaptive models, 120
adaptive parameters, 373
adaptive resonance, 118
Adaptive Resonance Theory I, 104, 107, 108, 115, 117, 243
algorithm for calculations, 246
equations for, 246
F1 layer calculations, 247
F2 layer calculations, 247
modifying connection weights, 248
Top−down inputs, 247
two−thirds rule, 244, 245
Adaptive Resonance Theory II, 248
Adaptive Resonance Theory III, 248
additive composition method, 506
advancing issues, 387
aggregation, 82, 87, 98
Ahmadian, 513
Aiken, 405
algorithm, 1
backpropagation, 7, 103, 271, 373, 375
constructive, 121
data clustering, 245
encoding, 94, 96
Index 429
learning algorithm , 61, 79, 102, 118

adaptive steepest descent, 410
alpha, 273, 274, 330, 372, 373, 384
alphabet classifier system, 320
Amari, 104
analog, 73, 74, 322
signal, 8
AND operation , 64
Anderson, 105
annealing
process, 430
schedule 113
simulated annealing, 113, 114
Anzai, 456, 472
application, 102, 374, 511
nature of , 74
approximation, 109
architecture, 7, 81, 120, 373
Artificial Intelligence , 6, 34
artificial neural network , 1
ART1 see Adaptive Resonance Theory I
ART2 see Adaptive Resonance Theory II
ART3 see Adaptive Resonance Theory III
artneuron class, 249
ASCII, 305, 306, 307, 329
graphic characters, 306
assocpair class
in BAM network, 186
association, 218
asynchronously , 1, 14
asynchronous update, 13, 62
attentional subsystem, 107, 243
Augusteijn, 512
autoassociation , 7, 8, 82,, 92, 102, 180
autoassociative network, 7, 64, 97, 375
average winner distance, 296
Azoff, 410
B
backpropagated errors, 144
Backpropagation, 10, 103, 104, 120, 123, 302, 325, 329, 374
algorithm, 7, 103, 271, 373, 375
beta, 330
calculating error, 396
calculations, 128, 130
changing beta while training, 337
choosing middle layer size, 372
convergence, 372
momentum term, 330
Index 430
noise factor, 336

self−supervised, 375
simulator, 134, 173, 337, 375, 377, 396
training and testing, 396
training law, 333
variations of , 373
Baer, 516
BAM see Bi−directional Associative Memory
bar
chart, 403, 404
features, 513
Barron’s, 377, 378, 388
base class , 25, 28, 138
beta, 136, 337, 372, 373, 384
bias, 16, 77, 125, 128, 378
Bi−directional Associative Memory, 81, 92, 104, 115, 117, 179, 185, 215
connection weight matrix, 212
continuous, 211
inputs, 180
network, 104
outputs, 180
training, 181
Unipolar Binary, 212
bin, 325
binary , 8, 15, 51, 65, 73, 74, 104
input pattern, 51, 98
patterns 11, 97
string, 16, 62
binary to bipolar mapping, 62, 63
binding
dynamic binding , 24, 139
late binding , 24
static binding, 139
bipolar, 15, 17, 104
mapping, 97
string, 16, 62, 180
bit pattern, 13
Blending problem, 418
block averages, 393
bmneuron class, 186
Boltzmann distribution, 113
Boltzmann machine, 92, 112, 113, 118, 419, 512
Booch, 21
boolean logic, 50
bottom−up
connection weight, 248
connections, 107, 244
Box−Jenkins methodology, 406
Brain−State−in−a−Box, 81, 82, 105
breadth, 387, 389
Buckles, 484
Index 431
buy/sell signals, 409, 410
C
cache, 137
Cader, 406
CAM see Content−Addressable−Memory
Carpenter, 92, 107, 117, 243, 269, 517
car navigation, 374
cartesian product, 479
cascade−correlation, 512
Cash Register game, 3, 65
categorization of inputs, 261
category, 37
Cauchy distribution, 113
Cauchy machine, 112, 113, 419
cells
complex cells, 106
simple cells, 106
center of area method, 504
centroid, 507, 508
character recognition, 305
characters
alphabetic, 305
ASCII, 306, 307
garbled, 322
graphic, 306, 307
handwritten, 320
Chawla, 514
Chiang, 513
cin, 25, 58, 71
clamping probabilities, 114
Clinton, Hillary, 405
C language, 21
class, 22
abstract class, 138
base class, 25, 28, 138, 139
derived class, 23, 25, 26, 144
friend class, 23
hierarchy, 27, 138, 139
input_layer class, 138
iostream class, 71
network class, 53, 66
output_layer class, 138
parent class, 26
classification, 322
C layer, 106
codebook vectors, 116
Cohen, 212
Collard, 405
Index 432
column vector, 97
combinatorial problem, 422
comparison layer, 244
competition, 9, 94, 97
competitive learning, 243
compilers
C++ compilers, 27
compilation error messages, 27
complement, 33, 185, 201, 202
complex cells, 106
composition
max−min, 220
compressor, 375
Computer Science, 34
conditional fuzzy
mean, 491
variance, 491
conjugate gradient methods, 373
conjunction operator, 505
connections, 2, 93, 102
bottom−up, 107
excitatory, 272
feedback , 82
inhibitory , 272
lateral, 93, 97, 107, 272, 276
recurrent, 82, 107, 179
top−down, 107
connection weight matrix, 220
connection weights, 89, 98
conscience factor, 302
constraints, 417
constructor, 23, 28, 55, 66
default constructor, 23
Consumer Price Index, 387
Content−Addressable−Memory, 5
continuous
Bi−directional Associative Memory, 211
models, 98
convergence, 78, 79, 96, 118, 119, 132, 323, 372, 373, 375
cooperation, 9, 94
correlation matrix, 9, 63
correlation−product encoding, 220
cost function, 124, 373
Cottrell, 374
counterpropagation, 106
network, 92, 93, 302
cout, 25, 58
C++, 21
classes, 138
code, 36
comments, 58
Index 433
compilers, 27
implementation, 185
crisp, 31, 73, 74
data sets, 475
rules, 48, 217
values, 50
cube, 84, 87, 89, 90
cum_deltas, 331
cycle, 78, 125
learning cycle, 103
cyclic information, 380
D
data biasing, 378
data hiding, 21, 22
data clustering, 109, 245
data completion, 102
data compression, 102, 302
Deboeck, 406
Davalo, 457
de la Maza, 410
Decision support systems, 75
declining issues, 387
decompressor, 375
decorrelation, 384
default
constructor, 23, 66
destructor, 23
defuzzification, 504, 506
degenerate tour, 424
degree of
certainty, 31
membership, 32, 477
degrees of freedom, 383
delete, 24, 144
delta rule, 110−113
derived class, 23, 25, 26, 144
descendant, 139, 143
DeSieno, 302
destructor, 23, 24
digital signal processing boards, 325
dimensionality, 381, 382, 384
directed graph, 65
discount rate, 35
discretization of a character, 98
discrete models, 98
discriminator, 517
disjunction operator, 506
display_input_char function, 308
Index 434
display_winner_weights function, 308

distance
Euclidean, 13
Hamming, 13
DJIA see Dow Jones Industrial Average
DOF see degrees of freedom
domains, 479, 484
dot product, 11, 12, 51, 64, 65
Dow Jones Industrial Average, 378, 386
dual confirmation trading system, 408
dynamic allocation of memory, 24
dynamic binding, 24, 139
dynamics
adaptive, 74
nonadaptive, 74
E
EMA see exponential moving average
encapsulate, 29
encapsulation, 21, 22
encode, 375
encoding, 7, 81, 220
algorithm, 94, 96
correlation−product, 220
phase, 94
thermometer, 380
time, 120
energy, 422
function, 119
level, 113, 422
surface, 177
Engineering, 34
epoch, 125
Ercal, 514
error signal, 103
error surface, 113
error tolerance, 136
Euclidean distance, 13, 280, 296
excitation, 94, 98, 276, 303
excitatory connections, 244, 272
exclusive or, 83
exemplar, 181−183, 201
class in BAM network, 186
pairs, 135, 177
patterns, 135, 177
exemplar pattern, 16, 64
exemplars, 64, 65, 74, 75, 115, 181
Expert systems, 48, 75, 217
exponential moving average, 399
Index 435
extended (database) model, 486

extended−delta−bar−delta, 406
F
factorial, 420
FAM see Fuzzy Associative Memories
Fast Fourier Transform, 374
fault tolerance, 374
feature detector, 328
feature extraction, 7, 513
Federal Reserve, 388, 388
feedback, 4, 5, 93
feed forward
Backpropagation, 81, 92, 112, 115, 123, 384,
network , 145, 406, 409, 511,
architecture, 124,
layout, 124,
network, 10
operation, 185
type, 2
field, 82
filter, 322
financial forecasting, 377
fire, 3, 71, 87, 89
first derivative, 380
fiscal policy, 36
fit values, 32, 217
fit vector, 32, 217, 221
floating point calculations, 519
compilation for, 519
F1 layer, 245
calculations, 247
Fogel, 76, 77
forecasting, 102
model, 378
T−Bill yields, 405
T−Note yields, 405
forward, 93, 104
Fourier transform, 380
Frank, 515
Freeman, 246, 248
frequency
component, 380
signature, 374
frequency spikes, 380
friend class, 23, 66, 68
F2 layer, 245
calculations, 247
Index 436
Fukushima, 92, 106

function
constructor function, 28
energy function, 119
evaluation, 109
fuzzy step function, 101
hyperbolic tangent function, 100
linear function, 99, 102
logistic function, 86, 100
Lyapunov function, 118, 119
member function, 28, 308
objective, 417
overloading, 25, 139
ramp function, 99, 101, 102
reference function, 493
sigmoid
function, 99, 100, 126, 129, 133, 164, 177, 395
logistic function, 100
step function, 99, 101
threshold function, 52, 95, 99, 101
XOR function, 83−85, 87
fuzzifier, 35, 36, 47
program, 50
fuzziness, 48, 50
fuzzy adaptive system, 49
fuzzy ARTMAP, 517
fuzzy association, 217
Fuzzy Associative Memories, 49, 50, 81, 92, 104, 115, 117, 217, 218, 473
encoding, 219, 220
Fuzzy Cognitive Map, 48, 49
fuzzy conditional expectations, 490, 509
fuzzy control , 497, 509
system, 47, 48, 473
fuzzy controller, 47
fuzzy database, 473, 475, 509
fuzzy expected value, 490
fuzzy equivalence relation, 481
fuzzy events, 488, 509
conditional probability of, 491
probability of , 490
fuzzy inputs, 47, 73
fuzzy logic, 31, 34, 50, 473
controller, 473, 497, 509
fuzzy matrix, 217
fuzzy means, 488, 490, 509
fuzzy numbers, 493, 496
triangular, 496
fuzzy OR method, 505
fuzzy outputs, 47, 74
fuzzy quantification, 488
fuzzy queries, 483, 488
Index 437
fuzzy relations, 479, 509

matrix representation, 479
fuzzy rule base, 502−504
fuzzy rules, 47, 50
fuzzy set, 32, 50, 218, 477, 488
complement, 218
height, 218
normal, 218
operations, 32, 217
fuzzy systems, 50, 217
fuzzy−valued, 34
fuzzy values, 477
fuzzy variances, 488, 490, 509
fzneuron class, 221
G
Gader, 513
gain , 107
constant, 273
parameter, 429, 467
gain control, 243, 248
unit, 244
Gallant, 117
Ganesh, 514
Gaussian density function, 458, 459, 524
generalization, 121, 320, 382
ability, 121, 336
generalized delta rule, 112, 176
genetic algorithms, 75, 76, 385
global
minimum, 113, 177
variable, 28
Glover, 471
gradient, 112, 113
grandmother cells, 117
gray scale, 305, 306, 322, 374
grid, 214, 305
Grossberg, 19, 92, 93, 107, 117, 243, 269
Grossberg layer, 9, 19, 82, 92, 106, 302
H
Hamming distance, 13, 201, 202
handwriting analysis, 98
handwriting recognition, 102
handwritten characters, 92
heap, 144
Hebb, 110
Hebbian
Index 438
conditioned learning, 302

learning, 105, 110
Hebb’s rule, 110, 111
Hecht−Nielsen, 93, 106, 302
Herrick Payoff Index, 401
heteroassociation, 7, 8, 82, 92, 102, 104, 180, 181
heteroassociative network, 7, 97
hidden layer, 2, 4, 75, 86, 89
hierarchical neural network, 407
hierarchies of classes, 27, 29
Hinton, 114
Hoff, 102, 112
holographic neural network, 408
Honig, 515
Hopfield, 422, 427, 429
memory, 73, 115, 117, 181
Hopfield network, 9, 11−14, 16, 19, 51, 79, 81, 82, 93, 111, 119, 120, 181, 472
Hotelling transform, 384
Housing Starts, 387
hybrid models, 75
hyperbolic tangent function, 429
hypercube, 218
unit, 218
hyperplane, 84
hypersphere, 273
I
image, 106, 302
compression, 374
processing, 98, 102
five−way transform, 516
recognition, 375
resolution, 322
implementation of functions, 67
ineuron class, 66
inference engine, 47
inheritance, 21, 25, 26, 138
multiple inheritance, 26
inhibition, 9, 94, 98
global, 428, 456
lateral, 272, 276
inhibitory connection, 272
initialization of
bottom−up weights, 250
parameters, 246
top−down weights, 250
weights, 94
inline, 165
Index 439
input, 98
binary input, 98
bipolar input, 98
layer, 2, 10
nature of , 73
number of , 74
patterns, 51, 65
signals, 65
space, 124
vector, 53, 71, 272, 112
input/output, 71
inqreset function, 251
instar, 93
interactions, 94
interconnections, 7
interest rate, 387
internal activation , 3
intersection, 32, 33
inverse mapping, 62, 182
Investor’s Business Daily, 388
iostream, 54, 71
istream, 58
iterative process, 78
J
Jagota, 514
January effect, 380
Jurik, 381, 384, 392
K
Karhunen−Loev transform, 384
Katz, 377
Kimoto, 408
kohonen.dat file, 275, 298, 300, 317
Kohonen, 19, 116, 117, 245, 271, 303, 456
Kohonen feature map, 16, 271, 273, 303, 305, 323
conscience factor, 302
neighborhood size, 280, 299, 300
training law, 273
Kohonen layer, 9, 19, 82, 92, 106, 298, 302, 322
class, 276
Kohonen network, 275, 276, 280, 300, 303, 322
applications of, 302,
Kohonen output layer, 275
Kohonen Self−Organizing Map, 115, 456, 471, 472
Kosaka, 409,
Kosko, 49, 50, 104, 215, 242, 506
Kostenius, 408, 409
Index 440
Kronecker delta function 428, 524
L
lambda, 136, 433
late binding, 24
lateral, 93
lateral competition, 303
laterally connected, 65
lateral connections, 93, 97, 107, 272, 276
lateral inhibition, 272, 276
layer, 2, 81
C layer, 106
comparison, 244
complex layer, 106
F1, 244
F2, 244
Grossberg layer, 82, 92, 302
hidden layer, 75, 81, 86, 89
input layer, 2, 3, 82
Kohonen layer, 82, 92, 302, 322
middle layer, 329, 372
output layer, 2, 82
recognition, 244
S layer, 106
simple layer, 106
layout, 52, 86, 124
ART1, 244
BAM , 180
Brain−State−in−a−Box, 105
FAM, 219
Hopfield network, 11
for TSP, 427
LVQ, 117
Madaline model, 103
LBS Capital Management, 377
learning, 4, 74, 98, 109, 110, 117, 118
algorithm, 61, 79, 102, 118
cycle, 103
Hebbian, 105, 110
one−shot, 117
probabilistic, 113
rate(parameter), 111, 112, 123, 125, 127, 136, 175
supervised learning, 5, 110, 112, 117, 121
time, 120
unsupervised
learning, 5, 110, 117, 121
Learning Vector Quantizer, 115−117, 302
least mean squared error, 111, 119, 123, 419
Index 441
rule, 111
Le Cun, 375
Lee, 512
Levenberg−Marquardt optimization, 373
Lewis, 377
Lin, 512
linear function, 99, 102
linear possibility regression model, 493, 496, 509
linear programming, 417
integer, 417
linearly separable, 83− 85
LMS see least mean squared error rule
local minimum, 113, 177, 325
logic
boolean logic, 50
fuzzy logic, 31, 34, 50, 473
logical operations, 31
AND, 64
logistic function, 86, 100
Long−term memory, 6, 77− 79, 118, 472
traces of, 243
look−up
memory, 5
table, 106
LTM see Long−term memory
LVQ see Learning Vector Quantizer
Lyapunov Function, 118, 119
M
MACD see moving average convergence divergence
Madaline, 102, 103
main diagonal, 63, 480
malignant melanoma, 514
malloc, 24
Mandelman, 378
MAPE see mean absolute percentage error
mapping, 123, 180
binary to bipolar, 62, 63
inverse, 62, 182
nonlinear, 109
real to binary, 180
mapping surface, 109
Markowitz, 470
Marquez, 406
Mason, 516
matrix, 97, 521
addition, 521
correlation matrix, 9
fuzzy, 217
Index 442
multiplication, 11
product, 104, 522
transpose, 11
weight matrix, 97
max_cycles, 326
maximum, 33, 219
max−min composition, 220, 221
McClelland, 103
McCulloch, 6
McNeill, 508
mean absolute percentage error, 406
mean squared error, 111, 410
Mears, 212, 214
membership, 32
functions, 50
triangular, 499, 506
rules, 49
memorization, 121, 320, 336, 382, 397
memorizing inputs, 273, 320
memory, 98
adaptive, 471
fuzzy associative, 218, 221
long−term memory, 6, 77−79, 118, 472
recency based, 471
short−term memory, 6, 77, 78, 79, 107, 118, 471
Mendelsohn, 407, 408
methods, 22, 145
metric, 5, 103
mexican hat function, 272, 273, 274, 276
middle layer, 329, 372
choosing size 372
minimum, 33, 219
global, 113, 177
local, 113, 177, 325
Minsky, 83
model
ART1, 245
continuous, 98
discrete, 98
Perceptron model, 65, 68, 81, 83
modulo, 423
Mohamed, 513
momentum, 325, 330, 337, 372, 400
implementing, 331
parameter, 119, 123, 134, 384
term, 330, 375
Morse, 514
Moss, 514
moving average convergence divergence, 401
moving averages, 380, 399
simple, 399
Index 443
weighted, 399
multilayered, 92
network, 106
multilayer feed−forward neural network, 7
multilayer networks, 123
multiple inheritance, 26
Munro, 374
N
NF see noise factor
NP−complete, 419, 427, 457
NYSE see New York Stock Exchange
Naim, 457
neighborhood, 276, 303, 457
size, 274, 280, 299, 300
neighbors, 272
Neocognitron, 81, 92, 106
Nellis, 516
NETTalk, 374
network
adaptive, 77
architecture, 77, 384, 388
autoassociative network, 97
backpropagation network, 329
bi−directional associative memory network, 104,, 88
Brain−State−in−a−Box network, 105
class, 53, 66
heteroassociative networks, 97
Hopfield network, 9, 11−14, 16, 19, 51, 79, 81, 82, 93, 111, 119, 120, 181, 472
layout, 86
modeling, 73
multilayer network, 123
nodes, 65
Perceptron network, 65, 66, 68, 79
radial basis function networks, 112, 114, 115
RBF networks see radial basis function networks
self−organizing , 269
statistical trained networks, 112
NeuFuz, 49
neural network, 1, 2
algorithms, 176
artificial neural network, 1
autoassociative, 375
biological, 272
counterpropagation, 302
fault tolerance of, 374
FAM, 218
hierarchical, 407
holographic, 408
Index 444
Kohonen, 271, 322

multilayer, 123
Perceptron, 65
plug−in boards, 325
self−organizing, 107, 269
Tabu, 471
two−layer, 92
neural−trained fuzzy systems, 49
neuromimes 2
neuron, 1, 3
input neurons, 82
output neuron, 99
new, 24, 144
Newton’s method, 373
New York Stock Exchange, 387
new highs, 389
new lows, 389
noise, 5, 330, 336, 337, 372, 375
random, 375
noise factor, 336
noise−saturation dilemma, 79
noise tolerance, 105
noisy data, 320
nonlinear function, 120
nonlinear mapping, 109
nonlinear scaling function, 10
nonlinear optimization, 417, 422, 472
nontraining mode, 135
Normal distribution, 524
normal fuzzy set, 218
normalization of a vector, 272
normalized inputs, 271, 381
normalized weights, 280
notation, 132
nprm parameter, 433
number bins, 43
NYSE see New York Stock Exchange
O
object, 22
objective function, 417
object−orientation, 21
object−oriented programming language, 21
Object−Oriented Analysis and Design, 21
on−balance volume, 402
on center, off surround, 97, 98
one−shot learning, 117
oneuron class, 66
Operations Research, 34
Index 445
operator overloading, 25, 139

optical character
recognition, 245
recognizer, 514
optimization, 102, 109, 417
nonlinear, 417, 422, 472
stock portfolio, 470
ordered pair, 32
organization of layers for backpropagation program, 144
orienting subsystem, 107, 243
orthogonal, 11, 12, 51, 64, 98
bit patterns, 64
input vectors, 299
set, 65
ostream, 58
output, 99
layer, 2, 10
nature of , 74
number of , 74
space, 124
stream, 58
outstar, 93, 106
overfitting of data, 383
overlap composition method, 506
overloading, 21, 24
function overloading, 25, 139
operator overloading, 25, 139
overtrained network, 329
P
Papert, 83
Parker, 103
partitioning, 87, 88
past_deltas, 331
Patil, 406
pattern
association, 16
binary pattern, 11, 97
bit pattern, 99
character, 17
classification, 8, 98, 109
completion, 105, 109
matching, 8, 98
recognition, 16, 34, 102, 108, 305, 322
studies, 380
system, 121
spatial, 99, 214
Pavlovian, 5
Perceptron, 3, 4, 66, 73, 82, 87, 90, 93, 102, 112
Index 446
model, 65, 68, 81, 83

multilayer Perceptron, 85, 88
network, 65, 66, 68, 79
single−layer Perceptron, 85
permutations, 420
Perry, 484
perturbation , 5, 113
phoneme, 303, 374
phonetic typewriter, 303
Pimmel, 513
Pitts, 6
pixel, 16, 322, 329, 374
values, 214, 305
plastic, 107
plasticity, 78, 79
plasticity−stability dilemma, 243
Pletta, 515
polymorphic function, 24, 28
polymorphism, 21, 24, 27, 138
Pomerleau, 374
portfolio selection, 470, 472
possibility distributions, 486, 487, 509
relational model, 486, 487
postprocess, 35
postprocessing, 50
filter , 50
potlpair class
in BAM network, 186
preprocess, 35, 50
preprocessing , 87, 379, 399
data, 389
filter, 50
fuzzifier, 35
Price is Right, 3, 65
principal component analysis, 384
private, 23, 26
probabilities, 31, 419
probability, 31, 43
distributions, 113
processing
additive, 75
hybrid, 75
multiplicative, 75
PROJECT operation, 485
proposition, 31
protected, 23, 26, 54, 143
public, 23, 26, 53, 54
Index 447
quadratic form, 119, 120, 418

quadratic programming problem, 418
quantification, 473, 475
quantization levels, 322
queries, 475, 476, 488
fuzzy, 483, 488
R
radial basis function networks, 112, 114, 115
ramp function, 99, 101,, 102
Ramsay, 515
random number generator, 37
range, 394
normalized, 394, 395
rate of change, 392, 400, 404
function, 392
indicator, 393, 394
real−time recurrent learning algorithm, 515
recall, 7, 81, 184, 220
BAM , 184
FAM, 220, 221
recency based memory, 471
recognition layer, 244
recurrent, 13, 179
recurrent connections, 82, 107, 179
reference
activation level, 456
function, 493
regression analysis, 406
risk−adjusted return, 410
relations, 476
antisymmetric, 480, 481
reflexive, 480, 481
resemblance, 509
similarity, 481, 509
symmetric, 480, 481
transitive, 480, 481
relative strength index, 400, 404
remainder, 423
reset, 243, 247, 251, 262
reset node, 244
reset unit, 244
resonance, 104, 107, 117, 118, 181, 215, 243, 269
responsive exploration, 471
Ressler, 512
restrmax function, 251
return type, 23
reuse, 26
Robotics, 34
Index 448
ROC see rate of change

Rosenberg, 374
Ross, 517
row vector, 97, 104
RSI see relative strength index
rule
delta, 110, 111
generalized delta, 112
Hebbian, 111
Hebb’s, 110
rules
fuzzy rules, 50
Rumbaugh, 21
Rummelhart, 103
S
SP500 Index see Standard and Poor’s 500 Index
Sathyanarayan Rao, 515
saturate, 381
scalar, 61
scalar multiplier, 64
second derivative, 380
Sejnowski, 114, 374
self−adaptation, 115
self−connected, 53
self−organization, 5, 6, 74, 115, 116
self−organizing feature map, 116
Self−Organizing Map, 245, 271
self−organizing neural networks, 107, 117, 121
self−supervised backpropagation, 375
sensitivity analysis, 384
separable, 84, 86, 88
linearly separable, 83, 84
subsets, 87
separability, 84, 86
separating
line, 86
plane, 85
Sethuraman, 515
set membership, 32
Sharda, 406
shift operator, 25
Short Term Memory, 6, 77, 78, 79, 107, 118, 471
traces of, 243
Sigma Pi neural network , 75
sigmoid
activation function, 381, 387
function, 77, 99, 126, 129, 133, 164, 177, 395
squashing, 381
Index 449
signal filtering, 102

signals
analog, 98
similarity, 486
class, 481, 509
level, 486
relation, 481
simple cells, 106
simple moving average, 399
simulated annealing, 113, 114
simulator, 372, 396
controls, 173
mode, 138
Skapura, 246, 248
Slater, 515
S layer, 106
SMA see simple moving average
SOM see Self−Organizing Map
sonar target recognition, 374
spatial
pattern, 99, 214
temporal pattern, 105,
speech
recognition, 303,
synthesizer, 374,
spike, 380
squared error, 103
squashing
function, 384, 458, 459
stable, 79, 107
stability 78, 79, 118
and plasticity, 77
stability−plasticity dilemma, 79, 107,, 269
STM see Short Term Memory
Standard and Poor’s 500 Index, 377, 378
forecasting, 386
standard I/O routines, 519
state energy, 118
state machine, 48
static binding, 139
Steele, 514
steepest descent, 112, 113, 177, 373
step function, 99, 101
stochastics, 402, 404
Stoecker, 514
Stonham, 516
string
binary, 62
bipolar, 62
structure, 7, 7
subsample, 322
Index 450
subset, 221
subsystem
attentional, 107, 243
orienting, 107, 243
Sudjianto, 516
summand, 422
summation symbol, 422
supervised , 109
learning, 5, 110, 112, 115, 117, 121
training 94, 110, 115, 121, 125
Sweeney , 516
symbolic approach, 6
T
TSR see Terminate and Stay Resident
Tabu , 471
active, 471
neural network, 471
search, 471, 472
Tank, 422, 427, 429
target 378, 395
outputs, 110, 115
patterns, 105
scaled, 395
tau, 433
technical analysis, 399
temperature, 118
Temporal Associative Memory, 92
Terano, 496
Terminate and Stay Resident programs, 519
terminating value, 298
termination criterion, 322
test.dat file, 327, 328
test mode, 135, 137, 138, 164, 173, 327, 396
Thirty−year Treasury Bond Rate, 387
Three−month Treasury Bill Rate, 387
threshold
function, 2, 3, 12, 17, 19, 52, 95, 99, 101, 125, 183
value, 16, 52, 66, 77, 86, 87, 90, 101, 128, 456
thresholding, 87, 185
function, 133, 177, 182, 184, 214
Thro, 508
Tic−Tac−Toe, 76, 79
time lag, 380
time series forecasting, 406, 410
time shifting, 395
timeframe, 378
tolerance, 119, 125, 173, 245, 318, 322, 328, 329, 372
level, 78, 123
Index 451
value, 119
top−down
connection weight , 248
top−down inputs, 247
topology, 7
Topology Preserving Maps, 116
tour, 420
traces, 243
of STM, 243
of LTM, 243
trading
commodities, 405,
system, 378
dual confirmation, 408
training, 4, 74, 75, 98, 109, 110, 119, 181, 396
fast, 107
law, 272, 273, 274, 330, 333
mode, 135, 137, 138, 164, 173, 396
supervised, 94, 110, 115
slow, 107
time, 329
unsupervised, 107, 110
transpose
of a matrix, 11, 179, 181, 183
of a vector, 11, 63, 97, 181
traveling salesperson(salesman) problem, 118, 119, 419
hand calculation, 423
Hopfield network solution−Hopfield, 427
Hopfield network solution−Anzai, 456
Kohonen network solution, 456
triple, 217
truth value, 31
tsneuron class, 430
TS see Tabu search
TSP see traveling salesperson problem
turning point predictor, 409
turning points, 407
two−layer networks, 92
two−thirds rule, 107, 244, 245, 269
U
Umano, 486
Unemployment Rate, 387
undertrained network, 329
uniform distribution, 77
union, 32
Unipolar Binary Bi−directional Associative Memory, 212
unit
Index 452
circle, 299
hypercube, 218
unit length, 273
universal set, 33
universe of discourse, 498, 499
unsupervised , 107
learning, 5, 110, 115, 117, 121
training, 107, 110
V
value
fit value, 32
threshold value, 16, 52, 66, 77, 86, 87, 90, 101, 128, 456
variable
external, 28
global, 28
vector, 17
codebook vectors, 116
column vector, 97, 104, 181
fit vector, 32, 33
heterassociated, 181
input vector, 53, 71, 272, 112
normalization of, 272
potentially associated, 181
quantization, 302
row vector, 97, 181
weight vector, 9, 96
vector pairs, 181
vertex, 88
vertices, 88
vigilance parameter, 107, 243, 245, 247, 262
virtual, 24, 139
trading, 377
visibility, 26
visible, 53
W
walk−forward methodology, 408
Wall Street Journal, 388
Wasserman, 516
weight matrix, 9, 17, 51, 65, 97, 181, 183
weight sharing, 375
weight surface, 113
weight update, 276
weight vector, 9, 96
quantizing, 307
Index 453
weighted sum, 2, 3, 19, 271

weights , 4, 181
bottom−up, 250
connection , 89, 98
top−down, 250
Werbos, 103
Wetherell, 514
Widrow, 102, 112
winner indexes, 298
winner, 98
neuron, 323
winner−take−all, 97, 98, 115, 116, 243, 271, 274
World Wide Web, 388
Wu, 515
X
XOR function, 83−85, 87
Y
Yan, 473, 497
Yu, 212, 214
Yuret, 410
Z
zero pattern, 65
zero−one programming problem, 471
Zipser, 374
Table of Contents
Index 454

C++ Neural Networks & Fuzzy Logic - Valluru B Rao

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

C++ Neural Networks & Fuzzy Logic - Valluru B Rao

Uploaded by

Copyright:

Available Formats

C++ Neural Networks and Fuzzy Logic:Preface

C++ Neural Networks and Fuzzy Logic

The subjects are covered as follows:

Copyright © IDG Books Worldwide, Inc.

C++ Neural Networks and Fuzzy Logic

Chapter 1—Introduction to Neural Networks

Chapter 2—C++ and Object Orientation

Chapter 3—A Look at Fuzzy Logic

Chapter 4—Constructing a Neural Network

Chapter 5—A Survey of Neural Network Models

Neural Network Models

Chapter 6—Learning and Training

Chapter 8—BAM: Bidirectional Associative Memory

Chapter 9—FAM: Fuzzy Associative Memory

FAM Neural Network

Chapter 10—Adaptive Resonance Theory (ART)

Chapter 11—The Kohonen Self−Organizing Map

Chapter 12—Application to Pattern Recognition

Chapter 14—Application to Financial Forecasting

Storing Data in Different Files

Chapter 15—Application to Nonlinear Optimization

C++ Implementation of the Hopfield Network for the Traveling Salesperson

Chapter 16—Applications of Fuzzy Logic

Chapter 17—Further Applications

C++ Neural Networks and Fuzzy Logic

Vydehamma, Annapurnamma, Anandarao, Madanagopalarao, Govindarao, and Rajyalakshamma.

Copyright © IDG Books Worldwide, Inc.

C++ Neural Networks and Fuzzy Logic

Chapter 1 Introduction to Neural Networks 14

Figure 1.1 A typical neural network.

Previous Table of Contents Next

Copyright © IDG Books Worldwide, Inc.

C++ Neural Networks and Fuzzy Logic

Cash Register Game

Cash Register Game 16

Supervised or Unsupervised Learning

Previous Table of Contents Next

Copyright © IDG Books Worldwide, Inc.

C++ Neural Networks and Fuzzy Logic

Neural Network Construction

1. Structure—the architecture and topology of the neural network

Previous Table of Contents Next

Copyright © IDG Books Worldwide, Inc.

Neural Network Construction 19

C++ Neural Networks and Fuzzy Logic

Qualifying for a Mortgage

Cooperation and Competition

Previous Table of Contents Next

Copyright © IDG Books Worldwide, Inc.

C++ Neural Networks and Fuzzy Logic

Example—A Feed−Forward Network

Figure 1.2 A feed−forward neural network with topology 2−2−1.

For example, for x3 and x5:

Example—A Hopfield Network

Example—A Feed−Forward Network 22

Figure 1.3 Layout of a Hopfield network.

A1B1 + A2 B2 + A3B3 + A4B4 = (1x0 + 0x1 + 1x0 + 0x1) = 0

Previous Table of Contents Next

Copyright © IDG Books Worldwide, Inc.

Example—A Feed−Forward Network 23

C++ Neural Networks and Fuzzy Logic

[radic][(a – e)2 + (b – f)2 + (c – g)2 + (d – h)2 ]